X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fpinger.c;h=b479dc6f47f9733057f13c294254d0a8a97852fa;hp=b1995e6c6c30c800cb7364158511c6ed91962541;hb=917655fc2938b90a9c246dd2d58408c42aa1658d;hpb=b11d0d89b22e8409cac12129a80044f84df0023c diff --git a/lustre/ptlrpc/pinger.c b/lustre/ptlrpc/pinger.c index b1995e6..b479dc6 100644 --- a/lustre/ptlrpc/pinger.c +++ b/lustre/ptlrpc/pinger.c @@ -1,765 +1,691 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Portal-RPC reconnection and replay operations, for use in recovery. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * Copyright (c) 2003 Cluster File Systems, Inc. - * Authors: Phil Schwan - * Mike Shaver + * 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. * - * 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 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). * - * 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. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.gnu.org/licenses/gpl-2.0.html * - * 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. + * GPL HEADER END + */ +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. * - * 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. + * Copyright (c) 2011, 2015, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/ptlrpc/pinger.c + * + * Portal-RPC reconnection and replay operations, for use in recovery. */ -#ifndef __KERNEL__ -#include -#else #define DEBUG_SUBSYSTEM S_RPC -#endif +#include +#include #include #include #include "ptlrpc_internal.h" -struct semaphore pinger_sem; -static struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports); +static int suppress_pings; +module_param(suppress_pings, int, 0644); +MODULE_PARM_DESC(suppress_pings, "Suppress pings"); -int ptlrpc_ping(struct obd_import *imp) +struct mutex pinger_mutex; +static struct list_head pinger_imports = + LIST_HEAD_INIT(pinger_imports); +static struct list_head timeout_list = + LIST_HEAD_INIT(timeout_list); + +int ptlrpc_pinger_suppress_pings(void) { - struct ptlrpc_request *req; - int rc = 0; - ENTRY; - - req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL); - if (req) { - DEBUG_REQ(D_INFO, req, "pinging %s->%s", - imp->imp_obd->obd_uuid.uuid, - obd2cli_tgt(imp->imp_obd)); - req->rq_no_resend = req->rq_no_delay = 1; - ptlrpc_req_set_repsize(req, 1, NULL); - req->rq_timeout = PING_INTERVAL; - ptlrpcd_add_req(req); - } else { - CERROR("OOM trying to ping %s->%s\n", - imp->imp_obd->obd_uuid.uuid, - obd2cli_tgt(imp->imp_obd)); - rc = -ENOMEM; - } - - RETURN(rc); + return suppress_pings; } +EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings); -static void ptlrpc_update_next_ping(struct obd_import *imp) +struct ptlrpc_request * +ptlrpc_prep_ping(struct obd_import *imp) { - imp->imp_next_ping = cfs_time_shift( - (imp->imp_state == LUSTRE_IMP_DISCON ? - RECONNECT_INTERVAL : PING_INTERVAL)); + struct ptlrpc_request *req; + + req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, + LUSTRE_OBD_VERSION, OBD_PING); + if (req) { + ptlrpc_request_set_replen(req); + req->rq_no_resend = req->rq_no_delay = 1; + } + return req; } -void ptlrpc_ping_import_soon(struct obd_import *imp) +int ptlrpc_obd_ping(struct obd_device *obd) { - imp->imp_next_ping = cfs_time_current(); + int rc; + struct ptlrpc_request *req; + + ENTRY; + + req = ptlrpc_prep_ping(obd->u.cli.cl_import); + if (!req) + RETURN(-ENOMEM); + + req->rq_send_state = LUSTRE_IMP_FULL; + + rc = ptlrpc_queue_wait(req); + + ptlrpc_req_finished(req); + + RETURN(rc); } +EXPORT_SYMBOL(ptlrpc_obd_ping); -static inline int ptlrpc_next_reconnect(struct obd_import *imp) +static bool ptlrpc_check_import_is_idle(struct obd_import *imp) { - if (imp->imp_server_timeout) - return cfs_time_shift(obd_timeout / 2); - else - return cfs_time_shift(obd_timeout); + struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; + time64_t now; + + if (!imp->imp_idle_timeout) + return false; + + /* + * 4 comes from: + * - client_obd_setup() - hashed import + * - ptlrpcd_alloc_work() + * - ptlrpcd_alloc_work() + * - ptlrpc_pinger_add_import + */ + if (atomic_read(&imp->imp_refcount) > 4) + return false; + + /* any lock increases ns_bref being a resource holder */ + if (ns && atomic_read(&ns->ns_bref) > 0) + return false; + + now = ktime_get_real_seconds(); + if (now - imp->imp_last_reply_time < imp->imp_idle_timeout) + return false; + + return true; } -static atomic_t suspend_timeouts = ATOMIC_INIT(0); -static cfs_time_t suspend_wakeup_time = 0; +static int ptlrpc_ping(struct obd_import *imp) +{ + struct ptlrpc_request *req; + + ENTRY; -#ifdef __KERNEL__ -static wait_queue_head_t suspend_timeouts_waitq; -#endif + if (ptlrpc_check_import_is_idle(imp)) + RETURN(ptlrpc_disconnect_and_idle_import(imp)); -cfs_time_t ptlrpc_suspend_wakeup_time(void) + req = ptlrpc_prep_ping(imp); + if (!req) { + CERROR("OOM trying to ping %s->%s\n", + imp->imp_obd->obd_uuid.uuid, + obd2cli_tgt(imp->imp_obd)); + RETURN(-ENOMEM); + } + + DEBUG_REQ(D_INFO, req, "pinging %s->%s", + imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); + ptlrpcd_add_req(req); + + RETURN(0); +} + +static void ptlrpc_update_next_ping(struct obd_import *imp, int soon) { - return suspend_wakeup_time; +#ifdef ENABLE_PINGER + time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL; + + if (imp->imp_state == LUSTRE_IMP_DISCON) { + time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN, + AT_OFF ? 0 : + at_get(&imp->imp_at.iat_net_latency)); + time = min(time, dtime); + } + imp->imp_next_ping = ktime_get_seconds() + time; +#endif /* ENABLE_PINGER */ } -void ptlrpc_deactivate_timeouts(struct obd_import *imp) +void ptlrpc_ping_import_soon(struct obd_import *imp) { - /*XXX: disabled for now, will be replaced by adaptive timeouts */ -#if 0 - if (imp->imp_no_timeout) - return; - imp->imp_no_timeout = 1; - atomic_inc(&suspend_timeouts); - CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts)); -#endif + imp->imp_next_ping = ktime_get_seconds(); } -void ptlrpc_activate_timeouts(struct obd_import *imp) +static inline int imp_is_deactive(struct obd_import *imp) { - /*XXX: disabled for now, will be replaced by adaptive timeouts */ -#if 0 - if (!imp->imp_no_timeout) - return; - imp->imp_no_timeout = 0; - LASSERT(atomic_read(&suspend_timeouts) > 0); - if (atomic_dec_and_test(&suspend_timeouts)) { - suspend_wakeup_time = cfs_time_current(); - wake_up(&suspend_timeouts_waitq); - } - CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts)); -#endif + return imp->imp_deactive || + OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE); } -int ptlrpc_check_suspend(void) +static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp) { - if (atomic_read(&suspend_timeouts)) - return 1; - return 0; + if (imp->imp_server_timeout) + return ktime_get_seconds() + (obd_timeout >> 1); + else + return ktime_get_seconds() + obd_timeout; } -int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req) +static time64_t pinger_check_timeout(time64_t time) { - struct l_wait_info lwi; - - if (atomic_read(&suspend_timeouts)) { - DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout", - atomic_read(&suspend_timeouts)); - lwi = LWI_INTR(NULL, NULL); - l_wait_event(suspend_timeouts_waitq, - atomic_read(&suspend_timeouts) == 0, &lwi); - DEBUG_REQ(D_NET, req, "-- recharge regular timeout"); - return 1; - } - return 0; + struct timeout_item *item; + time64_t timeout = PING_INTERVAL; + + /* This list is sorted in increasing timeout order */ + mutex_lock(&pinger_mutex); + list_for_each_entry(item, &timeout_list, ti_chain) { + time64_t ti_timeout = item->ti_timeout; + + if (timeout > ti_timeout) + timeout = ti_timeout; + break; + } + mutex_unlock(&pinger_mutex); + + return time + timeout - ktime_get_seconds(); } -#ifdef __KERNEL__ +static bool ir_up; -static void ptlrpc_pinger_process_import(struct obd_import *imp, - unsigned long this_ping) +void ptlrpc_pinger_ir_up(void) { - int force, level; - - spin_lock(&imp->imp_lock); - level = imp->imp_state; - force = imp->imp_force_verify; - if (force) - imp->imp_force_verify = 0; - spin_unlock(&imp->imp_lock); - - CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, - "level %s/%u force %u deactive %u pingable %u\n", - ptlrpc_import_state_name(level), level, - force, imp->imp_deactive, imp->imp_pingable); - - if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, - this_ping) && force == 0) - return; - - if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) { - /* wait at least a timeout before trying recovery again */ - imp->imp_next_ping = ptlrpc_next_reconnect(imp); - ptlrpc_initiate_recovery(imp); - } else if (level != LUSTRE_IMP_FULL || - imp->imp_obd->obd_no_recov || - imp->imp_deactive) { - CDEBUG(D_HA, "not pinging %s (in recovery " - " or recovery disabled: %s)\n", - obd2cli_tgt(imp->imp_obd), - ptlrpc_import_state_name(level)); - } else if (imp->imp_pingable || force) { - ptlrpc_ping(imp); - } + CDEBUG(D_HA, "IR up\n"); + ir_up = true; } +EXPORT_SYMBOL(ptlrpc_pinger_ir_up); -static int ptlrpc_pinger_main(void *arg) +void ptlrpc_pinger_ir_down(void) { - struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg; - struct ptlrpc_thread *thread = data->thread; - ENTRY; - - cfs_daemonize(data->name); - - /* Record that the thread is running */ - thread->t_flags = SVC_RUNNING; - cfs_waitq_signal(&thread->t_ctl_waitq); - - /* And now, loop forever, pinging as needed. */ - while (1) { - cfs_time_t this_ping = cfs_time_current(); - struct l_wait_info lwi; - cfs_duration_t time_to_next_ping; - struct list_head *iter; - - mutex_down(&pinger_sem); - list_for_each(iter, &pinger_imports) { - struct obd_import *imp = - list_entry(iter, struct obd_import, - imp_pinger_chain); - - ptlrpc_pinger_process_import(imp, this_ping); - /* obd_timeout might have changed */ - if (imp->imp_pingable && imp->imp_next_ping && - cfs_time_after(imp->imp_next_ping, - cfs_time_add(this_ping, - cfs_time_seconds(PING_INTERVAL)))) - ptlrpc_update_next_ping(imp); - } - mutex_up(&pinger_sem); - - /* Wait until the next ping time, or until we're stopped. */ - time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping, - cfs_time_seconds(PING_INTERVAL)), - cfs_time_current()); - - /* The ping sent by ptlrpc_send_rpc may get sent out - say .01 second after this. - ptlrpc_pinger_eending_on_import will then set the - next ping time to next_ping + .01 sec, which means - we will SKIP the next ping at next_ping, and the - ping will get sent 2 timeouts from now! Beware. */ - CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n", - time_to_next_ping, - cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL))); - if (time_to_next_ping > 0) { - lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_ping, cfs_time_seconds(1)), - NULL, NULL); - l_wait_event(thread->t_ctl_waitq, - thread->t_flags & (SVC_STOPPING|SVC_EVENT), - &lwi); - if (thread->t_flags & SVC_STOPPING) { - thread->t_flags &= ~SVC_STOPPING; - EXIT; - break; - } else if (thread->t_flags & SVC_EVENT) { - /* woken after adding import to reset timer */ - thread->t_flags &= ~SVC_EVENT; - } - } - } - - thread->t_flags = SVC_STOPPED; - cfs_waitq_signal(&thread->t_ctl_waitq); - - CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid()); - return 0; + CDEBUG(D_HA, "IR down\n"); + ir_up = false; } +EXPORT_SYMBOL(ptlrpc_pinger_ir_down); -static struct ptlrpc_thread *pinger_thread = NULL; +static void ptlrpc_pinger_process_import(struct obd_import *imp, + time64_t this_ping) +{ + int level; + int force; + int force_next; + int suppress; + + spin_lock(&imp->imp_lock); + + level = imp->imp_state; + force = imp->imp_force_verify; + force_next = imp->imp_force_next_verify; + /* + * This will be used below only if the import is "FULL". + */ + suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS); + + imp->imp_force_verify = 0; + + if (imp->imp_next_ping - 5 >= this_ping && !force) { + spin_unlock(&imp->imp_lock); + return; + } + + imp->imp_force_next_verify = 0; + + CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, + "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n", + imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd), + ptlrpc_import_state_name(level), level, force, force_next, + imp->imp_deactive, imp->imp_pingable, suppress); + + if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) { + /* wait for a while before trying recovery again */ + imp->imp_next_ping = ptlrpc_next_reconnect(imp); + spin_unlock(&imp->imp_lock); + if (!imp->imp_no_pinger_recover || + imp->imp_connect_error == -EAGAIN) + ptlrpc_initiate_recovery(imp); + } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov || + imp_is_deactive(imp)) { + CDEBUG(D_HA, + "%s->%s: not pinging (in recovery or recovery disabled: %s)\n", + imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd), + ptlrpc_import_state_name(level)); + if (force) + imp->imp_force_verify = 1; + spin_unlock(&imp->imp_lock); + } else if ((imp->imp_pingable && !suppress) || force_next || force) { + spin_unlock(&imp->imp_lock); + ptlrpc_ping(imp); + } else { + spin_unlock(&imp->imp_lock); + } +} + +static struct workqueue_struct *pinger_wq; +static void ptlrpc_pinger_main(struct work_struct *ws); +static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main); + +static void ptlrpc_pinger_main(struct work_struct *ws) +{ + time64_t this_ping = ktime_get_seconds(); + time64_t time_to_next_wake; + struct timeout_item *item; + struct obd_import *imp; + struct list_head *iter; + + do { + mutex_lock(&pinger_mutex); + list_for_each_entry(item, &timeout_list, ti_chain) + item->ti_cb(item, item->ti_cb_data); + + list_for_each(iter, &pinger_imports) { + imp = list_entry(iter, struct obd_import, + imp_pinger_chain); + + ptlrpc_pinger_process_import(imp, this_ping); + /* obd_timeout might have changed */ + if (imp->imp_pingable && imp->imp_next_ping && + imp->imp_next_ping > this_ping + PING_INTERVAL) + ptlrpc_update_next_ping(imp, 0); + } + mutex_unlock(&pinger_mutex); + /* update memory usage info */ + obd_update_maxusage(); + + /* Wait until the next ping time, or until we're stopped. */ + time_to_next_wake = pinger_check_timeout(this_ping); + /* + * The ping sent by ptlrpc_send_rpc may get sent out + * say .01 second after this. + * ptlrpc_pinger_sending_on_import will then set the + * next ping time to next_ping + .01 sec, which means + * we will SKIP the next ping at next_ping, and the + * ping will get sent 2 timeouts from now! Beware. + */ + CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n", + time_to_next_wake, this_ping + PING_INTERVAL); + } while (time_to_next_wake <= 0); + + queue_delayed_work(pinger_wq, &ping_work, + cfs_time_seconds(max(time_to_next_wake, 1LL))); +} int ptlrpc_start_pinger(void) { - struct l_wait_info lwi = { 0 }; - struct ptlrpc_svc_data d; - int rc; -#ifndef ENABLE_PINGER - return 0; +#ifdef ENABLE_PINGER + if (pinger_wq) + return -EALREADY; + + pinger_wq = alloc_workqueue("ptlrpc_pinger", 0, 1); + if (!pinger_wq) { + CERROR("cannot start pinger workqueue\n"); + return -ENOMEM; + } + + queue_delayed_work(pinger_wq, &ping_work, 0); + + if (suppress_pings) + CWARN("Pings will be suppressed at the request of the administrator. The configuration shall meet the additional requirements described in the manual. (Search for the \"suppress_pings\" kernel module parameter.)\n"); #endif - ENTRY; - - if (pinger_thread != NULL) - RETURN(-EALREADY); - - OBD_ALLOC_PTR(pinger_thread); - if (pinger_thread == NULL) - RETURN(-ENOMEM); - cfs_waitq_init(&pinger_thread->t_ctl_waitq); - cfs_waitq_init(&suspend_timeouts_waitq); - - d.name = "ll_ping"; - d.thread = pinger_thread; - - /* 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_pinger_main, &d, CLONE_VM | CLONE_FILES); - if (rc < 0) { - CERROR("cannot start thread: %d\n", rc); - OBD_FREE(pinger_thread, sizeof(*pinger_thread)); - pinger_thread = NULL; - RETURN(rc); - } - l_wait_event(pinger_thread->t_ctl_waitq, - pinger_thread->t_flags & SVC_RUNNING, &lwi); - - RETURN(0); + return 0; } +int ptlrpc_pinger_remove_timeouts(void); + int ptlrpc_stop_pinger(void) { - struct l_wait_info lwi = { 0 }; - int rc = 0; -#ifndef ENABLE_PINGER - return 0; -#endif - ENTRY; - - if (pinger_thread == NULL) - RETURN(-EALREADY); - mutex_down(&pinger_sem); - pinger_thread->t_flags = SVC_STOPPING; - cfs_waitq_signal(&pinger_thread->t_ctl_waitq); - mutex_up(&pinger_sem); +#ifdef ENABLE_PINGER + if (!pinger_wq) + return -EALREADY; - l_wait_event(pinger_thread->t_ctl_waitq, - (pinger_thread->t_flags & SVC_STOPPED), &lwi); + ptlrpc_pinger_remove_timeouts(); - OBD_FREE_PTR(pinger_thread); - pinger_thread = NULL; - RETURN(rc); + cancel_delayed_work_sync(&ping_work); + destroy_workqueue(pinger_wq); + pinger_wq = NULL; +#endif + return 0; } void ptlrpc_pinger_sending_on_import(struct obd_import *imp) { - ptlrpc_update_next_ping(imp); + ptlrpc_update_next_ping(imp, 0); +} + +void ptlrpc_pinger_commit_expected(struct obd_import *imp) +{ + ptlrpc_update_next_ping(imp, 1); + assert_spin_locked(&imp->imp_lock); + /* + * Avoid reading stale imp_connect_data. When not sure if pings are + * expected or not on next connection, we assume they are not and force + * one anyway to guarantee the chance of updating + * imp_peer_committed_transno. + */ + if (imp->imp_state != LUSTRE_IMP_FULL || + OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS)) + imp->imp_force_next_verify = 1; } int ptlrpc_pinger_add_import(struct obd_import *imp) { - ENTRY; - if (!list_empty(&imp->imp_pinger_chain)) - RETURN(-EALREADY); - - mutex_down(&pinger_sem); - CDEBUG(D_HA, "adding pingable import %s->%s\n", - imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); - /* if we add to pinger we want recovery on this import */ - imp->imp_obd->obd_no_recov = 0; - ptlrpc_update_next_ping(imp); - /* XXX sort, blah blah */ - list_add_tail(&imp->imp_pinger_chain, &pinger_imports); - class_import_get(imp); - - ptlrpc_pinger_wake_up(); - mutex_up(&pinger_sem); - - RETURN(0); + ENTRY; + if (!list_empty(&imp->imp_pinger_chain)) + RETURN(-EALREADY); + + mutex_lock(&pinger_mutex); + CDEBUG(D_HA, "adding pingable import %s->%s\n", + imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); + /* if we add to pinger we want recovery on this import */ + imp->imp_obd->obd_no_recov = 0; + ptlrpc_update_next_ping(imp, 0); + /* XXX sort, blah blah */ + list_add_tail(&imp->imp_pinger_chain, &pinger_imports); + class_import_get(imp); + + ptlrpc_pinger_wake_up(); + mutex_unlock(&pinger_mutex); + + RETURN(0); } +EXPORT_SYMBOL(ptlrpc_pinger_add_import); int ptlrpc_pinger_del_import(struct obd_import *imp) { - ENTRY; - if (list_empty(&imp->imp_pinger_chain)) - RETURN(-ENOENT); - - mutex_down(&pinger_sem); - list_del_init(&imp->imp_pinger_chain); - CDEBUG(D_HA, "removing pingable import %s->%s\n", - imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); - /* if we remove from pinger we don't want recovery on this import */ - imp->imp_obd->obd_no_recov = 1; - class_import_put(imp); - mutex_up(&pinger_sem); - RETURN(0); + ENTRY; + + if (list_empty(&imp->imp_pinger_chain)) + RETURN(-ENOENT); + + mutex_lock(&pinger_mutex); + list_del_init(&imp->imp_pinger_chain); + CDEBUG(D_HA, "removing pingable import %s->%s\n", + imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); + /* if we remove from pinger we don't want recovery on this import */ + imp->imp_obd->obd_no_recov = 1; + class_import_put(imp); + mutex_unlock(&pinger_mutex); + RETURN(0); } +EXPORT_SYMBOL(ptlrpc_pinger_del_import); -void ptlrpc_pinger_wake_up() +/** + * Register a timeout callback to the pinger list, and the callback will + * be called when timeout happens. + */ +static struct timeout_item *ptlrpc_new_timeout(time64_t time, + enum timeout_event event, + timeout_cb_t cb, void *data) { -#ifdef ENABLE_PINGER - pinger_thread->t_flags |= SVC_EVENT; - cfs_waitq_signal(&pinger_thread->t_ctl_waitq); -#endif -} + struct timeout_item *ti; -/* Ping evictor thread */ -#define PET_READY 1 -#define PET_TERMINATE 2 + OBD_ALLOC_PTR(ti); + if (!ti) + return NULL; -static int pet_refcount = 0; -static int pet_state; -static wait_queue_head_t pet_waitq; -static struct obd_export *pet_exp = NULL; -static spinlock_t pet_lock = SPIN_LOCK_UNLOCKED; + INIT_LIST_HEAD(&ti->ti_obd_list); + INIT_LIST_HEAD(&ti->ti_chain); + ti->ti_timeout = time; + ti->ti_event = event; + ti->ti_cb = cb; + ti->ti_cb_data = data; -int ping_evictor_wake(struct obd_export *exp) -{ - spin_lock(&pet_lock); - if (pet_exp) { - /* eventually the new obd will call here again. */ - spin_unlock(&pet_lock); - return 1; - } - - /* We have to make sure the obd isn't destroyed between now and when - * the ping evictor runs. We'll take a reference here, and drop it - * when we finish in the evictor. We don't really care about this - * export in particular; we just need one to keep the obd alive. */ - pet_exp = class_export_get(exp); - spin_unlock(&pet_lock); - - wake_up(&pet_waitq); - return 0; + return ti; } -static int ping_evictor_main(void *arg) +/** + * Register timeout event on the the pinger thread. + * Note: the timeout list is an sorted list with increased timeout value. + */ +static struct timeout_item* +ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event, + timeout_cb_t cb, void *data) { - struct obd_device *obd; - struct obd_export *exp; - struct l_wait_info lwi = { 0 }; - time_t expire_time; - ENTRY; - - ptlrpc_daemonize("ll_evictor"); - - CDEBUG(D_HA, "Starting Ping Evictor\n"); - pet_exp = NULL; - pet_state = PET_READY; - while (1) { - l_wait_event(pet_waitq, pet_exp || - (pet_state == PET_TERMINATE), &lwi); - if (pet_state == PET_TERMINATE) - break; - - /* we only get here if pet_exp != NULL, and the end of this - * loop is the only place which sets it NULL again, so lock - * is not strictly necessary. */ - spin_lock(&pet_lock); - obd = pet_exp->exp_obd; - spin_unlock(&pet_lock); - - expire_time = CURRENT_SECONDS - (3 * obd_timeout / 2); - - CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n", - obd->obd_name, expire_time); - - /* Exports can't be deleted out of the list while we hold - * the obd lock (class_unlink_export), which means we can't - * lose the last ref on the export. If they've already been - * removed from the list, we won't find them here. */ - spin_lock(&obd->obd_dev_lock); - while (!list_empty(&obd->obd_exports_timed)) { - exp = list_entry(obd->obd_exports_timed.next, - struct obd_export,exp_obd_chain_timed); - if (expire_time > exp->exp_last_request_time) { - class_export_get(exp); - spin_unlock(&obd->obd_dev_lock); - LCONSOLE_WARN("%s: haven't heard from client %s" - " (at %s) in %ld seconds. I think" - " it's dead, and I am evicting" - " it. exp %p, cur %ld expire %ld" - " last %ld\n", - obd->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), - (long)(CURRENT_SECONDS - - exp->exp_last_request_time), - exp, (long)CURRENT_SECONDS, - (long)expire_time, - (long)exp->exp_last_request_time); - CDEBUG(D_HA, "Last request was at %ld\n", - exp->exp_last_request_time); - class_fail_export(exp); - class_export_put(exp); - spin_lock(&obd->obd_dev_lock); - } else { - /* List is sorted, so everyone below is ok */ - break; - } - } - spin_unlock(&obd->obd_dev_lock); - - class_export_put(pet_exp); - - spin_lock(&pet_lock); - pet_exp = NULL; - spin_unlock(&pet_lock); - } - CDEBUG(D_HA, "Exiting Ping Evictor\n"); - - RETURN(0); + struct timeout_item *item, *tmp; + + LASSERT(mutex_is_locked(&pinger_mutex)); + + list_for_each_entry(item, &timeout_list, ti_chain) + if (item->ti_event == event) + goto out; + + item = ptlrpc_new_timeout(time, event, cb, data); + if (item) { + list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) { + if (tmp->ti_timeout < time) { + list_add(&item->ti_chain, &tmp->ti_chain); + goto out; + } + } + list_add(&item->ti_chain, &timeout_list); + } +out: + return item; } -void ping_evictor_start(void) +/* Add a client_obd to the timeout event list, when timeout(@time) + * happens, the callback(@cb) will be called. + */ +int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event, + timeout_cb_t cb, void *data, + struct list_head *obd_list) { - int rc; - - if (++pet_refcount > 1) - return; - - init_waitqueue_head(&pet_waitq); - - rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES); - if (rc < 0) { - pet_refcount--; - CERROR("Cannot start ping evictor thread: %d\n", rc); - } + struct timeout_item *ti; + + mutex_lock(&pinger_mutex); + ti = ptlrpc_pinger_register_timeout(time, event, cb, data); + if (!ti) { + mutex_unlock(&pinger_mutex); + return -EINVAL; + } + list_add(obd_list, &ti->ti_obd_list); + mutex_unlock(&pinger_mutex); + return 0; } -EXPORT_SYMBOL(ping_evictor_start); +EXPORT_SYMBOL(ptlrpc_add_timeout_client); -void ping_evictor_stop(void) +int ptlrpc_del_timeout_client(struct list_head *obd_list, + enum timeout_event event) { - if (--pet_refcount > 0) - return; - - pet_state = PET_TERMINATE; - wake_up(&pet_waitq); + struct timeout_item *ti = NULL, *item; + + if (list_empty(obd_list)) + return 0; + mutex_lock(&pinger_mutex); + list_del_init(obd_list); + /** + * If there are no obd attached to the timeout event + * list, remove this timeout event from the pinger + */ + list_for_each_entry(item, &timeout_list, ti_chain) { + if (item->ti_event == event) { + ti = item; + break; + } + } + LASSERTF(ti, "ti is NULL !\n"); + if (list_empty(&ti->ti_obd_list)) { + list_del(&ti->ti_chain); + OBD_FREE_PTR(ti); + } + mutex_unlock(&pinger_mutex); + return 0; } -EXPORT_SYMBOL(ping_evictor_stop); -#else /* !__KERNEL__ */ - -/* XXX - * the current implementation of pinger in liblustre is not optimized - */ +EXPORT_SYMBOL(ptlrpc_del_timeout_client); -#ifdef ENABLE_PINGER -static struct pinger_data { - int pd_recursion; - cfs_time_t pd_this_ping; /* jiffies */ - cfs_time_t pd_next_ping; /* jiffies */ - struct ptlrpc_request_set *pd_set; -} pinger_args; - -static int pinger_check_rpcs(void *arg) +int ptlrpc_pinger_remove_timeouts(void) { - cfs_time_t curtime = cfs_time_current(); - struct ptlrpc_request *req; - struct ptlrpc_request_set *set; - struct list_head *iter; - struct pinger_data *pd = &pinger_args; - int rc; - - /* prevent recursion */ - if (pd->pd_recursion++) { - CDEBUG(D_HA, "pinger: recursion! quit\n"); - LASSERT(pd->pd_set); - pd->pd_recursion--; - return 0; - } - - /* have we reached ping point? */ - if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) { - pd->pd_recursion--; - return 0; - } - - /* if we have rpc_set already, continue processing it */ - if (pd->pd_set) { - LASSERT(pd->pd_this_ping); - set = pd->pd_set; - goto do_check_set; - } - - pd->pd_this_ping = curtime; - pd->pd_set = ptlrpc_prep_set(); - if (pd->pd_set == NULL) - goto out; - set = pd->pd_set; - - /* add rpcs into set */ - mutex_down(&pinger_sem); - list_for_each(iter, &pinger_imports) { - struct obd_import *imp = - list_entry(iter, struct obd_import, imp_pinger_chain); - int generation, level; - - if (cfs_time_aftereq(pd->pd_this_ping, - imp->imp_next_ping - 5 * CFS_TICK)) { - /* Add a ping. */ - spin_lock(&imp->imp_lock); - generation = imp->imp_generation; - level = imp->imp_state; - spin_unlock(&imp->imp_lock); - - if (level != LUSTRE_IMP_FULL) { - CDEBUG(D_HA, - "not pinging %s (in recovery)\n", - obd2cli_tgt(imp->imp_obd)); - continue; - } - - req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, - 1, NULL, NULL); - if (!req) { - CERROR("out of memory\n"); - break; - } - req->rq_no_resend = 1; - ptlrpc_req_set_repsize(req, 1, NULL); - req->rq_send_state = LUSTRE_IMP_FULL; - req->rq_phase = RQ_PHASE_RPC; - req->rq_import_generation = generation; - ptlrpc_set_add_req(set, req); - } else { - CDEBUG(D_HA, "don't need to ping %s ("CFS_TIME_T" > " - CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd), - imp->imp_next_ping, pd->pd_this_ping); - } - } - pd->pd_this_ping = curtime; - mutex_up(&pinger_sem); - - /* Might be empty, that's OK. */ - if (set->set_remaining == 0) - CDEBUG(D_HA, "nothing to ping\n"); - - list_for_each(iter, &set->set_requests) { - struct ptlrpc_request *req = - list_entry(iter, struct ptlrpc_request, - rq_set_chain); - DEBUG_REQ(D_HA, req, "pinging %s->%s", - req->rq_import->imp_obd->obd_uuid.uuid, - obd2cli_tgt(req->rq_import->imp_obd)); - (void)ptl_send_rpc(req, 0); - } - -do_check_set: - rc = ptlrpc_check_set(set); - - /* not finished, and we are not expired, simply return */ - if (!rc && cfs_time_before(curtime, - cfs_time_add(pd->pd_this_ping, - cfs_time_seconds(PING_INTERVAL)))) { - CDEBUG(D_HA, "not finished, but also not expired\n"); - pd->pd_recursion--; - return 0; - } - - /* Expire all the requests that didn't come back. */ - mutex_down(&pinger_sem); - list_for_each(iter, &set->set_requests) { - req = list_entry(iter, struct ptlrpc_request, - rq_set_chain); - - if (req->rq_phase == RQ_PHASE_COMPLETE) - continue; - - req->rq_phase = RQ_PHASE_COMPLETE; - atomic_dec(&req->rq_import->imp_inflight); - set->set_remaining--; - /* If it was disconnected, don't sweat it. */ - if (list_empty(&req->rq_import->imp_pinger_chain)) { - ptlrpc_unregister_reply(req); - continue; - } - - CDEBUG(D_HA, "pinger initiate expire_one_request\n"); - ptlrpc_expire_one_request(req); - } - mutex_up(&pinger_sem); - - ptlrpc_set_destroy(set); - pd->pd_set = NULL; - -out: - pd->pd_next_ping = cfs_time_add(pd->pd_this_ping, - cfs_time_seconds(PING_INTERVAL)); - pd->pd_this_ping = 0; /* XXX for debug */ - - CDEBUG(D_HA, "finished a round ping\n"); - pd->pd_recursion--; - return 0; + struct timeout_item *item, *tmp; + + mutex_lock(&pinger_mutex); + list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) { + LASSERT(list_empty(&item->ti_obd_list)); + list_del(&item->ti_chain); + OBD_FREE_PTR(item); + } + mutex_unlock(&pinger_mutex); + return 0; } -static void *pinger_callback = NULL; -#endif /* ENABLE_PINGER */ - -int ptlrpc_start_pinger(void) +void ptlrpc_pinger_wake_up(void) { #ifdef ENABLE_PINGER - memset(&pinger_args, 0, sizeof(pinger_args)); - pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs", - &pinger_check_rpcs, - &pinger_args); + mod_delayed_work(pinger_wq, &ping_work, 0); #endif - return 0; } -int ptlrpc_stop_pinger(void) +/* Ping evictor thread */ +#define PET_READY 1 +#define PET_TERMINATE 2 + +static int pet_refcount; +static int pet_state; +static wait_queue_head_t pet_waitq; +static struct list_head pet_list; +static DEFINE_SPINLOCK(pet_lock); + +int ping_evictor_wake(struct obd_export *exp) { -#ifdef ENABLE_PINGER - if (pinger_callback) - liblustre_deregister_wait_callback(pinger_callback); -#endif - return 0; + struct obd_device *obd; + + spin_lock(&pet_lock); + if (pet_state != PET_READY) { + /* eventually the new obd will call here again. */ + spin_unlock(&pet_lock); + return 1; + } + + obd = class_exp2obd(exp); + if (list_empty(&obd->obd_evict_list)) { + class_incref(obd, "evictor", obd); + list_add(&obd->obd_evict_list, &pet_list); + } + spin_unlock(&pet_lock); + + wake_up(&pet_waitq); + return 0; } -void ptlrpc_pinger_sending_on_import(struct obd_import *imp) +static int ping_evictor_main(void *arg) { -#ifdef ENABLE_PINGER - mutex_down(&pinger_sem); - ptlrpc_update_next_ping(imp); - if (pinger_args.pd_set == NULL && - time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) { - CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n", - imp->imp_next_ping, cfs_time_current()); - pinger_args.pd_next_ping = imp->imp_next_ping; - } - mutex_up(&pinger_sem); -#endif + struct obd_device *obd; + struct obd_export *exp; + time64_t expire_time; + + ENTRY; + unshare_fs_struct(); + + CDEBUG(D_HA, "Starting Ping Evictor\n"); + pet_state = PET_READY; + while (1) { + wait_event_idle(pet_waitq, + (!list_empty(&pet_list)) || + (pet_state == PET_TERMINATE)); + + /* loop until all obd's will be removed */ + if ((pet_state == PET_TERMINATE) && list_empty(&pet_list)) + break; + + /* + * we only get here if pet_exp != NULL, and the end of this + * loop is the only place which sets it NULL again, so lock + * is not strictly necessary. + */ + spin_lock(&pet_lock); + obd = list_entry(pet_list.next, struct obd_device, + obd_evict_list); + spin_unlock(&pet_lock); + + expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT; + + CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n", + obd->obd_name, expire_time); + + /* + * Exports can't be deleted out of the list while we hold + * the obd lock (class_unlink_export), which means we can't + * lose the last ref on the export. If they've already been + * removed from the list, we won't find them here. + */ + spin_lock(&obd->obd_dev_lock); + while (!list_empty(&obd->obd_exports_timed)) { + exp = list_entry(obd->obd_exports_timed.next, + struct obd_export, + exp_obd_chain_timed); + if (expire_time > exp->exp_last_request_time) { + struct obd_uuid *client_uuid; + + class_export_get(exp); + client_uuid = &exp->exp_client_uuid; + spin_unlock(&obd->obd_dev_lock); + LCONSOLE_WARN("%s: haven't heard from client %s (at %s) in %lld seconds. I think it's dead, and I am evicting it. exp %p, cur %lld expire %lld last %lld\n", + obd->obd_name, + obd_uuid2str(client_uuid), + obd_export_nid2str(exp), + ktime_get_real_seconds() - + exp->exp_last_request_time, + exp, ktime_get_real_seconds(), + expire_time, + exp->exp_last_request_time); + CDEBUG(D_HA, "Last request was at %lld\n", + exp->exp_last_request_time); + class_fail_export(exp); + class_export_put(exp); + spin_lock(&obd->obd_dev_lock); + } else { + /* List is sorted, so everyone below is ok */ + break; + } + } + spin_unlock(&obd->obd_dev_lock); + + spin_lock(&pet_lock); + list_del_init(&obd->obd_evict_list); + spin_unlock(&pet_lock); + + class_decref(obd, "evictor", obd); + } + CDEBUG(D_HA, "Exiting Ping Evictor\n"); + + RETURN(0); } -int ptlrpc_pinger_add_import(struct obd_import *imp) +void ping_evictor_start(void) { - ENTRY; - if (!list_empty(&imp->imp_pinger_chain)) - RETURN(-EALREADY); + struct task_struct *task; - CDEBUG(D_HA, "adding pingable import %s->%s\n", - imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); - ptlrpc_pinger_sending_on_import(imp); + if (++pet_refcount > 1) + return; - mutex_down(&pinger_sem); - list_add_tail(&imp->imp_pinger_chain, &pinger_imports); - class_import_get(imp); - mutex_up(&pinger_sem); + INIT_LIST_HEAD(&pet_list); + init_waitqueue_head(&pet_waitq); - RETURN(0); + task = kthread_run(ping_evictor_main, NULL, "ll_evictor"); + if (IS_ERR(task)) { + pet_refcount--; + CERROR("Cannot start ping evictor thread: %ld\n", + PTR_ERR(task)); + } } +EXPORT_SYMBOL(ping_evictor_start); -int ptlrpc_pinger_del_import(struct obd_import *imp) +void ping_evictor_stop(void) { - ENTRY; - if (list_empty(&imp->imp_pinger_chain)) - RETURN(-ENOENT); - - mutex_down(&pinger_sem); - list_del_init(&imp->imp_pinger_chain); - CDEBUG(D_HA, "removing pingable import %s->%s\n", - imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd)); - class_import_put(imp); - mutex_up(&pinger_sem); - RETURN(0); -} + if (--pet_refcount > 0) + return; -void ptlrpc_pinger_wake_up() -{ -#ifdef ENABLE_PINGER - /* XXX force pinger to run, if needed */ - struct obd_import *imp; - ENTRY; - list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) { - CDEBUG(D_HA, "Checking that we need to do anything about import" - " %s->%s\n", imp->imp_obd->obd_uuid.uuid, - obd2cli_tgt(imp->imp_obd)); -#ifdef ENABLE_LIBLUSTRE_RECOVERY - if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive) -#else - /*XXX only recover for the initial connection */ - if (!lustre_handle_is_used(&imp->imp_remote_handle) && - imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive) -#endif - ptlrpc_initiate_recovery(imp); - else if (imp->imp_state != LUSTRE_IMP_FULL) - CDEBUG(D_HA, "Refused to recover import %s->%s " - "state %d, deactive %d\n", - imp->imp_obd->obd_uuid.uuid, - obd2cli_tgt(imp->imp_obd), imp->imp_state, - imp->imp_deactive); - } - EXIT; -#endif + pet_state = PET_TERMINATE; + wake_up(&pet_waitq); } -#endif /* !__KERNEL__ */ +EXPORT_SYMBOL(ping_evictor_stop);