X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Frecov_thread.c;h=8076420dce4c20461e388f76607f18a82a0a884a;hp=8b386ee4e6277a4668cada70bec02f4e83283fc1;hb=6ad1581bf6003631d87e8365c01c03b06be136de;hpb=fbb7ead129258897f5a5d5c9ce28d31fbbe5bca2 diff --git a/lustre/ptlrpc/recov_thread.c b/lustre/ptlrpc/recov_thread.c index 8b386ee..8076420 100644 --- a/lustre/ptlrpc/recov_thread.c +++ b/lustre/ptlrpc/recov_thread.c @@ -1,605 +1,775 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Copyright (C) 2003 Cluster File Systems, Inc. - * Author: Andreas Dilger + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This file is part of Lustre, http://www.lustre.org. + * 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. * - * Lustre is free 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. + * 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). * - * 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 - * GNU General Public License for more details. + * 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 * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * 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. * - * OST<->MDS recovery logging thread. + * GPL HEADER END + */ +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, Whamcloud, Inc. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. * + * lustre/ptlrpc/recov_thread.c + * + * OST<->MDS recovery logging thread. * Invariants in implementation: * - we do not share logs among different OST<->MDS connections, so that * if an OST or MDS fails it need only look at log(s) relevant to itself + * + * Author: Andreas Dilger + * Yury Umanets + * Alexey Lyashkov */ #define DEBUG_SUBSYSTEM S_LOG -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif - #ifdef __KERNEL__ -#include +# include #else # include # include #endif -#include -#include -#include -#include -#include -#include -#include +#include +#include +#include +#include +#include #include -#include +#include #include "ptlrpc_internal.h" -#define LLCD_SIZE 4096 +static cfs_atomic_t llcd_count = CFS_ATOMIC_INIT(0); +static cfs_mem_cache_t *llcd_cache = NULL; #ifdef __KERNEL__ +enum { + LLOG_LCM_FL_START = 1 << 0, + LLOG_LCM_FL_EXIT = 1 << 1 +}; -static struct llog_commit_master lustre_lcm; -static struct llog_commit_master *lcm = &lustre_lcm; +struct llcd_async_args { + struct llog_canceld_ctxt *la_ctxt; +}; -/* Allocate new commit structs in case we do not have enough */ -static int llcd_alloc(void) +static void llcd_print(struct llog_canceld_ctxt *llcd, + const char *func, int line) { - struct llog_canceld_ctxt *llcd; - int offset = offsetof(struct llog_canceld_ctxt, llcd_cookies); + CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line); + CDEBUG(D_RPCTRACE, " size: %d\n", llcd->llcd_size); + CDEBUG(D_RPCTRACE, " ctxt: %p\n", llcd->llcd_ctxt); + CDEBUG(D_RPCTRACE, " lcm : %p\n", llcd->llcd_lcm); + CDEBUG(D_RPCTRACE, " cookiebytes : %d\n", llcd->llcd_cookiebytes); +} - OBD_ALLOC(llcd, LLCD_SIZE + offset); - if (llcd == NULL) - return -ENOMEM; +/** + * Allocate new llcd from cache, init it and return to caller. + * Bumps number of objects allocated. + */ +static struct llog_canceld_ctxt *llcd_alloc(struct llog_commit_master *lcm) +{ + struct llog_canceld_ctxt *llcd; + int size, overhead; + + LASSERT(lcm != NULL); + + /* + * We want to send one page of cookies with rpc header. This buffer + * will be assigned later to the rpc, this is why we preserve the + * space for rpc header. + */ + size = CFS_PAGE_SIZE - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL); + overhead = offsetof(struct llog_canceld_ctxt, llcd_cookies); + OBD_SLAB_ALLOC_GFP(llcd, llcd_cache, size + overhead, CFS_ALLOC_STD); + if (!llcd) + return NULL; + + CFS_INIT_LIST_HEAD(&llcd->llcd_list); + llcd->llcd_cookiebytes = 0; + llcd->llcd_size = size; + cfs_spin_lock(&lcm->lcm_lock); llcd->llcd_lcm = lcm; + cfs_atomic_inc(&lcm->lcm_count); + cfs_list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds); + cfs_spin_unlock(&lcm->lcm_lock); + cfs_atomic_inc(&llcd_count); - spin_lock(&lcm->lcm_llcd_lock); - list_add(&llcd->llcd_list, &lcm->lcm_llcd_free); - atomic_inc(&lcm->lcm_llcd_numfree); - spin_unlock(&lcm->lcm_llcd_lock); + CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n", + llcd, lcm, cfs_atomic_read(&lcm->lcm_count)); - return 0; + return llcd; } -/* Get a free cookie struct from the list */ -struct llog_canceld_ctxt *llcd_grab(void) +/** + * Returns passed llcd to cache. + */ +static void llcd_free(struct llog_canceld_ctxt *llcd) { - struct llog_canceld_ctxt *llcd; - - spin_lock(&lcm->lcm_llcd_lock); - if (list_empty(&lcm->lcm_llcd_free)) { - spin_unlock(&lcm->lcm_llcd_lock); - if (llcd_alloc() < 0) { - CERROR("unable to allocate log commit data!\n"); - return NULL; + struct llog_commit_master *lcm = llcd->llcd_lcm; + int size; + + if (lcm) { + if (cfs_atomic_read(&lcm->lcm_count) == 0) { + CERROR("Invalid llcd free %p\n", llcd); + llcd_print(llcd, __FUNCTION__, __LINE__); + LBUG(); } - spin_lock(&lcm->lcm_llcd_lock); + cfs_spin_lock(&lcm->lcm_lock); + LASSERT(!cfs_list_empty(&llcd->llcd_list)); + cfs_list_del_init(&llcd->llcd_list); + cfs_atomic_dec(&lcm->lcm_count); + cfs_spin_unlock(&lcm->lcm_lock); + + CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n", + llcd, lcm, cfs_atomic_read(&lcm->lcm_count)); } - llcd = list_entry(lcm->lcm_llcd_free.next, typeof(*llcd), llcd_list); - list_del(&llcd->llcd_list); - atomic_dec(&lcm->lcm_llcd_numfree); - spin_unlock(&lcm->lcm_llcd_lock); - - llcd->llcd_tries = 0; - llcd->llcd_cookiebytes = 0; + LASSERT(cfs_atomic_read(&llcd_count) > 0); + cfs_atomic_dec(&llcd_count); - return llcd; + size = offsetof(struct llog_canceld_ctxt, llcd_cookies) + + llcd->llcd_size; + OBD_SLAB_FREE(llcd, llcd_cache, size); } -EXPORT_SYMBOL(llcd_grab); -static void llcd_put(struct llog_canceld_ctxt *llcd) +/** + * Checks if passed cookie fits into llcd free space buffer. Returns + * 1 if yes and 0 otherwise. + */ +static inline int +llcd_fit(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies) { - int offset = offsetof(struct llog_canceld_ctxt, llcd_cookies); + return (llcd->llcd_size - llcd->llcd_cookiebytes >= sizeof(*cookies)); +} - if (atomic_read(&lcm->lcm_llcd_numfree) >= lcm->lcm_llcd_maxfree) { - OBD_FREE(llcd, LLCD_SIZE + offset); - } else { - spin_lock(&lcm->lcm_llcd_lock); - list_add(&llcd->llcd_list, &lcm->lcm_llcd_free); - atomic_inc(&lcm->lcm_llcd_numfree); - spin_unlock(&lcm->lcm_llcd_lock); - } +/** + * Copy passed @cookies to @llcd. + */ +static inline void +llcd_copy(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies) +{ + LASSERT(llcd_fit(llcd, cookies)); + memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes, + cookies, sizeof(*cookies)); + llcd->llcd_cookiebytes += sizeof(*cookies); } -/* Send some cookies to the appropriate target */ -void llcd_send(struct llog_canceld_ctxt *llcd) +/** + * Llcd completion function. Called uppon llcd send finish regardless + * sending result. Error is passed in @rc. Note, that this will be called + * in cleanup time when all inflight rpcs aborted. + */ +static int +llcd_interpret(const struct lu_env *env, + struct ptlrpc_request *req, void *args, int rc) { - spin_lock(&llcd->llcd_lcm->lcm_llcd_lock); - list_add_tail(&llcd->llcd_list, &llcd->llcd_lcm->lcm_llcd_pending); - spin_unlock(&llcd->llcd_lcm->lcm_llcd_lock); + struct llcd_async_args *la = args; + struct llog_canceld_ctxt *llcd = la->la_ctxt; - wake_up_nr(&llcd->llcd_lcm->lcm_waitq, 1); + CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc); + llcd_free(llcd); + return 0; } -EXPORT_SYMBOL(llcd_send); -/* deleted objects have a commit callback that cancels the MDS - * log record for the deletion. The commit callback calls this - * function +/** + * Send @llcd to remote node. Free llcd uppon completion or error. Sending + * is performed in async style so this function will return asap without + * blocking. */ -int llog_obd_repl_cancel(struct llog_ctxt *ctxt, int count, - struct llog_cookie *cookies, int flags, void *data) +static int llcd_send(struct llog_canceld_ctxt *llcd) { - struct llog_canceld_ctxt *llcd; - int rc = 0; + char *bufs[2] = { NULL, (char *)llcd->llcd_cookies }; + struct obd_import *import = NULL; + struct llog_commit_master *lcm; + struct llcd_async_args *la; + struct ptlrpc_request *req; + struct llog_ctxt *ctxt; + int rc; ENTRY; - LASSERT(ctxt); - down(&ctxt->loc_sem); - if (ctxt->loc_imp == NULL) { - CWARN("no import for ctxt %p\n", ctxt); - GOTO(out, rc = 0); + ctxt = llcd->llcd_ctxt; + if (!ctxt) { + CERROR("Invalid llcd with NULL ctxt found (%p)\n", + llcd); + llcd_print(llcd, __FUNCTION__, __LINE__); + LBUG(); } + LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex); - llcd = ctxt->loc_llcd; - if (count > 0 && cookies != NULL) { - if (llcd == NULL) { - llcd = llcd_grab(); - if (llcd == NULL) { - CERROR("couldn't get an llcd - dropped "LPX64":%x+%u\n", - cookies->lgc_lgl.lgl_oid, - cookies->lgc_lgl.lgl_ogen, cookies->lgc_index); - GOTO(out, rc = -ENOMEM); - } - llcd->llcd_ctxt = ctxt; - ctxt->loc_llcd = llcd; - } - memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes, - cookies, sizeof(*cookies)); - llcd->llcd_cookiebytes += sizeof(*cookies); - } else { - if (llcd == NULL || !(flags & OBD_LLOG_FL_SENDNOW)) - GOTO(out, rc); + if (llcd->llcd_cookiebytes == 0) + GOTO(exit, rc = 0); + + lcm = llcd->llcd_lcm; + + /* + * Check if we're in exit stage. Do not send llcd in + * this case. + */ + if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) + GOTO(exit, rc = -ENODEV); + + CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd); + + import = llcd->llcd_ctxt->loc_imp; + if (!import || (import == LP_POISON) || + (import->imp_client == LP_POISON)) { + CERROR("Invalid import %p for llcd %p\n", + import, llcd); + GOTO(exit, rc = -ENODEV); } - if ((LLCD_SIZE - llcd->llcd_cookiebytes < sizeof(*cookies) || - flags & OBD_LLOG_FL_SENDNOW)) { - CDEBUG(D_HA, "send llcd %p:%p\n", llcd, llcd->llcd_ctxt); - ctxt->loc_llcd = NULL; - llcd_send(llcd); + OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10); + + /* + * No need to get import here as it is already done in + * llog_receptor_accept(). + */ + req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL); + if (req == NULL) { + CERROR("Can't allocate request for sending llcd %p\n", + llcd); + GOTO(exit, rc = -ENOMEM); } -out: - up(&ctxt->loc_sem); + req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, + RCL_CLIENT, llcd->llcd_cookiebytes); + + rc = ptlrpc_request_bufs_pack(req, LUSTRE_LOG_VERSION, + OBD_LOG_CANCEL, bufs, NULL); + if (rc) { + ptlrpc_request_free(req); + GOTO(exit, rc); + } + + ptlrpc_at_set_req_timeout(req); + ptlrpc_request_set_replen(req); + + /* bug 5515 */ + req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL; + req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL; + + req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret; + + CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args)); + la = ptlrpc_req_async_args(req); + la->la_ctxt = llcd; + + /* llog cancels will be replayed after reconnect so this will do twice + * first from replay llog, second for resended rpc */ + req->rq_no_delay = req->rq_no_resend = 1; + + ptlrpc_set_add_new_req(&lcm->lcm_pc, req); + RETURN(0); +exit: + CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd); + llcd_free(llcd); return rc; } -EXPORT_SYMBOL(llog_obd_repl_cancel); -int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp) +/** + * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection + * so hat they can refer each other. + */ +static int +llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd) { - int rc = 0; - ENTRY; + LASSERT(ctxt != NULL && llcd != NULL); + LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex); + LASSERT(ctxt->loc_llcd == NULL); + llcd->llcd_ctxt = llog_ctxt_get(ctxt); + ctxt->loc_llcd = llcd; - if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) { - down(&ctxt->loc_sem); - if (ctxt->loc_llcd != NULL) { - llcd_put(ctxt->loc_llcd); - ctxt->loc_llcd = NULL; - } - CWARN("reverse import disconnected, put " - "llcd %p:%p\n", ctxt->loc_llcd, ctxt); - ctxt->loc_imp = NULL; - up(&ctxt->loc_sem); - } else { - rc = llog_cancel(ctxt, 0, NULL, OBD_LLOG_FL_SENDNOW, NULL); - } + CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n", + llcd, ctxt); - RETURN(rc); + return 0; } -EXPORT_SYMBOL(llog_obd_repl_sync); -static int log_commit_thread(void *arg) +/** + * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes + * sure that this llcd will not be found another time we try to cancel. + */ +static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt) { - struct llog_commit_master *lcm = arg; - struct llog_commit_daemon *lcd; - struct llog_canceld_ctxt *llcd, *n; - unsigned long flags; - ENTRY; - - OBD_ALLOC(lcd, sizeof(*lcd)); - if (lcd == NULL) - RETURN(-ENOMEM); + struct llog_canceld_ctxt *llcd; - lock_kernel(); - ptlrpc_daemonize(); /* thread never needs to do IO */ - - SIGNAL_MASK_LOCK(current, flags); - sigfillset(¤t->blocked); - RECALC_SIGPENDING; - SIGNAL_MASK_UNLOCK(current, flags); - - spin_lock(&lcm->lcm_thread_lock); - - THREAD_NAME(current->comm, sizeof(current->comm) - 1, - "ll_log_comt_%02d", atomic_read(&lcm->lcm_thread_total)); - atomic_inc(&lcm->lcm_thread_total); - spin_unlock(&lcm->lcm_thread_lock); - unlock_kernel(); - - INIT_LIST_HEAD(&lcd->lcd_lcm_list); - INIT_LIST_HEAD(&lcd->lcd_llcd_list); - lcd->lcd_lcm = lcm; - - CDEBUG(D_HA, "%s started\n", current->comm); - do { - struct ptlrpc_request *request; - struct obd_import *import = NULL; - struct list_head *sending_list; - int rc = 0; - - /* If we do not have enough pages available, allocate some */ - while (atomic_read(&lcm->lcm_llcd_numfree) < - lcm->lcm_llcd_minfree) { - if (llcd_alloc() < 0) - break; - } + LASSERT(ctxt != NULL); + LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex); - spin_lock(&lcm->lcm_thread_lock); - atomic_inc(&lcm->lcm_thread_numidle); - list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_idle); - spin_unlock(&lcm->lcm_thread_lock); - - wait_event_interruptible(lcm->lcm_waitq, - !list_empty(&lcm->lcm_llcd_pending) || - lcm->lcm_flags & LLOG_LCM_FL_EXIT); - - /* If we are the last available thread, start a new one in case - * we get blocked on an RPC (nobody else will start a new one)*/ - spin_lock(&lcm->lcm_thread_lock); - atomic_dec(&lcm->lcm_thread_numidle); - list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_busy); - spin_unlock(&lcm->lcm_thread_lock); - - sending_list = &lcm->lcm_llcd_pending; - resend: - if (lcm->lcm_flags & LLOG_LCM_FL_EXIT) { - lcm->lcm_llcd_maxfree = 0; - lcm->lcm_llcd_minfree = 0; - lcm->lcm_thread_max = 0; - - if (list_empty(&lcm->lcm_llcd_pending) || - lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE) - break; - } + llcd = ctxt->loc_llcd; + if (!llcd) + return NULL; - if (atomic_read(&lcm->lcm_thread_numidle) <= 1 && - atomic_read(&lcm->lcm_thread_total) < lcm->lcm_thread_max) { - rc = llog_start_commit_thread(); - if (rc < 0) - CERROR("error starting thread: rc %d\n", rc); - } + CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n", + llcd, ctxt); - /* Move all of the pending cancels from the same OST off of - * the list, so we don't get multiple threads blocked and/or - * doing upcalls on the same OST in case of failure. */ - spin_lock(&lcm->lcm_llcd_lock); - if (!list_empty(sending_list)) { - list_move_tail(sending_list->next, - &lcd->lcd_llcd_list); - llcd = list_entry(lcd->lcd_llcd_list.next, - typeof(*llcd), llcd_list); - LASSERT(llcd->llcd_lcm == lcm); - import = llcd->llcd_ctxt->loc_imp; - } - list_for_each_entry_safe(llcd, n, sending_list, llcd_list) { - LASSERT(llcd->llcd_lcm == lcm); - if (import == llcd->llcd_ctxt->loc_imp) - list_move_tail(&llcd->llcd_list, - &lcd->lcd_llcd_list); - } - if (sending_list != &lcm->lcm_llcd_resend) { - list_for_each_entry_safe(llcd, n, &lcm->lcm_llcd_resend, - llcd_list) { - LASSERT(llcd->llcd_lcm == lcm); - if (import == llcd->llcd_ctxt->loc_imp) - list_move_tail(&llcd->llcd_list, - &lcd->lcd_llcd_list); - } - } - spin_unlock(&lcm->lcm_llcd_lock); - - /* We are the only one manipulating our local list - no lock */ - list_for_each_entry_safe(llcd,n, &lcd->lcd_llcd_list,llcd_list){ - char *bufs[1] = {(char *)llcd->llcd_cookies}; - - list_del(&llcd->llcd_list); - if (llcd->llcd_cookiebytes == 0) { - CDEBUG(D_HA, "put empty llcd %p:%p\n", - llcd, llcd->llcd_ctxt); - llcd_put(llcd); - continue; - } - - down(&llcd->llcd_ctxt->loc_sem); - if (llcd->llcd_ctxt->loc_imp == NULL) { - up(&llcd->llcd_ctxt->loc_sem); - CWARN("import will be destroyed, put " - "llcd %p:%p\n", llcd, llcd->llcd_ctxt); - llcd_put(llcd); - continue; - } - - up(&llcd->llcd_ctxt->loc_sem); - request = ptlrpc_prep_req(import, LUSTRE_LOG_VERSION, - OBD_LOG_CANCEL, 1, - &llcd->llcd_cookiebytes, - bufs); - - if (request == NULL) { - rc = -ENOMEM; - CERROR("error preparing commit: rc %d\n", rc); - - spin_lock(&lcm->lcm_llcd_lock); - list_splice(&lcd->lcd_llcd_list, - &lcm->lcm_llcd_resend); - INIT_LIST_HEAD(&lcd->lcd_llcd_list); - spin_unlock(&lcm->lcm_llcd_lock); - break; - } - - request->rq_replen = lustre_msg_size(0, NULL); - down(&llcd->llcd_ctxt->loc_sem); - if (llcd->llcd_ctxt->loc_imp == NULL) { - up(&llcd->llcd_ctxt->loc_sem); - CWARN("import will be destroyed, put " - "llcd %p:%p\n", llcd, llcd->llcd_ctxt); - llcd_put(llcd); - ptlrpc_req_finished(request); - continue; - } - up(&llcd->llcd_ctxt->loc_sem); - rc = ptlrpc_queue_wait(request); - ptlrpc_req_finished(request); - - /* If the RPC failed, we put this and the remaining - * messages onto the resend list for another time. */ - if (rc == 0) { - llcd_put(llcd); - continue; - } - -#if 0 /* FIXME just put llcd, not put it on resend list */ - spin_lock(&lcm->lcm_llcd_lock); - list_splice(&lcd->lcd_llcd_list, &lcm->lcm_llcd_resend); - if (++llcd->llcd_tries < 5) { - CERROR("commit %p failed on attempt %d: rc %d\n", - llcd, llcd->llcd_tries, rc); - - list_add_tail(&llcd->llcd_list, - &lcm->lcm_llcd_resend); - spin_unlock(&lcm->lcm_llcd_lock); - } else { - spin_unlock(&lcm->lcm_llcd_lock); -#endif - CERROR("commit %p:%p drop %d cookies: rc %d\n", - llcd, llcd->llcd_ctxt, - (int)(llcd->llcd_cookiebytes / - sizeof(*llcd->llcd_cookies)), rc); - llcd_put(llcd); -#if 0 - } - break; -#endif - } + ctxt->loc_llcd = NULL; + llog_ctxt_put(ctxt); + return llcd; +} - if (rc == 0) { - sending_list = &lcm->lcm_llcd_resend; - if (!list_empty(sending_list)) - goto resend; - } - } while(1); - - /* If we are force exiting, just drop all of the cookies. */ - if (lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE) { - spin_lock(&lcm->lcm_llcd_lock); - list_splice(&lcm->lcm_llcd_pending, &lcd->lcd_llcd_list); - list_splice(&lcm->lcm_llcd_resend, &lcd->lcd_llcd_list); - list_splice(&lcm->lcm_llcd_free, &lcd->lcd_llcd_list); - spin_unlock(&lcm->lcm_llcd_lock); - - list_for_each_entry_safe(llcd, n, &lcd->lcd_llcd_list,llcd_list) - llcd_put(llcd); +/** + * Return @llcd cached in @ctxt. Allocate new one if required. Attach it + * to ctxt so that it may be used for gathering cookies and sending. + */ +static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt) +{ + struct llog_canceld_ctxt *llcd; + LASSERT(ctxt); + llcd = llcd_alloc(ctxt->loc_lcm); + if (!llcd) { + CERROR("Can't alloc an llcd for ctxt %p\n", ctxt); + return NULL; } + llcd_attach(ctxt, llcd); + return llcd; +} - spin_lock(&lcm->lcm_thread_lock); - list_del(&lcd->lcd_lcm_list); - spin_unlock(&lcm->lcm_thread_lock); - OBD_FREE(lcd, sizeof(*lcd)); +/** + * Deatch llcd from its @ctxt. Free llcd. + */ +static void llcd_put(struct llog_ctxt *ctxt) +{ + struct llog_canceld_ctxt *llcd; - spin_lock(&lcm->lcm_thread_lock); - atomic_dec(&lcm->lcm_thread_total); - spin_unlock(&lcm->lcm_thread_lock); - wake_up(&lcm->lcm_waitq); + llcd = llcd_detach(ctxt); + if (llcd) + llcd_free(llcd); +} - CDEBUG(D_HA, "%s exiting\n", current->comm); - return 0; +/** + * Detach llcd from its @ctxt so that nobody will find it with try to + * re-use. Send llcd to remote node. + */ +static int llcd_push(struct llog_ctxt *ctxt) +{ + struct llog_canceld_ctxt *llcd; + int rc; + + /* + * Make sure that this llcd will not be sent again as we detach + * it from ctxt. + */ + llcd = llcd_detach(ctxt); + if (!llcd) { + CERROR("Invalid detached llcd found %p\n", llcd); + llcd_print(llcd, __FUNCTION__, __LINE__); + LBUG(); + } + + rc = llcd_send(llcd); + if (rc) + CERROR("Couldn't send llcd %p (%d)\n", llcd, rc); + return rc; } -int llog_start_commit_thread(void) +/** + * Start recovery thread which actually deals llcd sending. This + * is all ptlrpc standard thread based so there is not much of work + * to do. + */ +int llog_recov_thread_start(struct llog_commit_master *lcm) { int rc; ENTRY; - if (atomic_read(&lcm->lcm_thread_total) >= lcm->lcm_thread_max) - RETURN(0); - - rc = kernel_thread(log_commit_thread, lcm, CLONE_VM | CLONE_FILES); - if (rc < 0) { - CERROR("error starting thread #%d: %d\n", - atomic_read(&lcm->lcm_thread_total), rc); + rc = ptlrpcd_start(-1, 1, lcm->lcm_name, &lcm->lcm_pc); + if (rc) { + CERROR("Error %d while starting recovery thread %s\n", + rc, lcm->lcm_name); RETURN(rc); } - - RETURN(0); + RETURN(rc); } -EXPORT_SYMBOL(llog_start_commit_thread); - -static struct llog_process_args { - struct semaphore llpa_sem; - struct llog_ctxt *llpa_ctxt; - void *llpa_cb; - void *llpa_arg; -} llpa; +EXPORT_SYMBOL(llog_recov_thread_start); -int llog_init_commit_master(void) +/** + * Stop recovery thread. Complement to llog_recov_thread_start(). + */ +void llog_recov_thread_stop(struct llog_commit_master *lcm, int force) { - INIT_LIST_HEAD(&lcm->lcm_thread_busy); - INIT_LIST_HEAD(&lcm->lcm_thread_idle); - spin_lock_init(&lcm->lcm_thread_lock); - atomic_set(&lcm->lcm_thread_numidle, 0); - init_waitqueue_head(&lcm->lcm_waitq); - INIT_LIST_HEAD(&lcm->lcm_llcd_pending); - INIT_LIST_HEAD(&lcm->lcm_llcd_resend); - INIT_LIST_HEAD(&lcm->lcm_llcd_free); - spin_lock_init(&lcm->lcm_llcd_lock); - atomic_set(&lcm->lcm_llcd_numfree, 0); - lcm->lcm_llcd_minfree = 0; - lcm->lcm_thread_max = 5; - /* FIXME initialize semaphore for llog_process_args */ - sema_init(&llpa.llpa_sem, 1); - return 0; + ENTRY; + + /* + * Let all know that we're stopping. This will also make + * llcd_send() refuse any new llcds. + */ + cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags); + + /* + * Stop processing thread. No new rpcs will be accepted for + * for processing now. + */ + ptlrpcd_stop(&lcm->lcm_pc, force); + + /* + * By this point no alive inflight llcds should be left. Only + * those forgotten in sync may still be attached to ctxt. Let's + * print them. + */ + if (cfs_atomic_read(&lcm->lcm_count) != 0) { + struct llog_canceld_ctxt *llcd; + cfs_list_t *tmp; + + CERROR("Busy llcds found (%d) on lcm %p\n", + cfs_atomic_read(&lcm->lcm_count), lcm); + + cfs_spin_lock(&lcm->lcm_lock); + cfs_list_for_each(tmp, &lcm->lcm_llcds) { + llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt, + llcd_list); + llcd_print(llcd, __FUNCTION__, __LINE__); + } + cfs_spin_unlock(&lcm->lcm_lock); + + /* + * No point to go further with busy llcds at this point + * as this is clear bug. It might mean we got hanging + * rpc which holds import ref and this means we will not + * be able to cleanup anyways. + * + * Or we just missed to kill them when they were not + * attached to ctxt. In this case our slab will remind + * us about this a bit later. + */ + LBUG(); + } + EXIT; } +EXPORT_SYMBOL(llog_recov_thread_stop); -int llog_cleanup_commit_master(int force) +/** + * Initialize commit master structure and start recovery thread on it. + */ +struct llog_commit_master *llog_recov_thread_init(char *name) { - lcm->lcm_flags |= LLOG_LCM_FL_EXIT; - if (force) - lcm->lcm_flags |= LLOG_LCM_FL_EXIT_FORCE; - wake_up(&lcm->lcm_waitq); + struct llog_commit_master *lcm; + int rc; + ENTRY; - wait_event_interruptible(lcm->lcm_waitq, - atomic_read(&lcm->lcm_thread_total) == 0); - return 0; + OBD_ALLOC_PTR(lcm); + if (!lcm) + RETURN(NULL); + + /* + * Try to create threads with unique names. + */ + snprintf(lcm->lcm_name, sizeof(lcm->lcm_name), + "lcm_%s", name); + + cfs_atomic_set(&lcm->lcm_count, 0); + cfs_atomic_set(&lcm->lcm_refcount, 1); + cfs_spin_lock_init(&lcm->lcm_lock); + CFS_INIT_LIST_HEAD(&lcm->lcm_llcds); + rc = llog_recov_thread_start(lcm); + if (rc) { + CERROR("Can't start commit thread, rc %d\n", rc); + GOTO(out, rc); + } + RETURN(lcm); +out: + OBD_FREE_PTR(lcm); + return NULL; } -EXPORT_SYMBOL(llog_cleanup_commit_master); +EXPORT_SYMBOL(llog_recov_thread_init); -static int log_process_thread(void *args) +/** + * Finalize commit master and its recovery thread. + */ +void llog_recov_thread_fini(struct llog_commit_master *lcm, int force) { - struct llog_process_args *data = args; - struct llog_ctxt *ctxt = data->llpa_ctxt; - void *cb = data->llpa_cb; - struct llog_logid logid = *(struct llog_logid *)(data->llpa_arg); - struct llog_handle *llh = NULL; - unsigned long flags; + ENTRY; + llog_recov_thread_stop(lcm, force); + lcm_put(lcm); + EXIT; +} +EXPORT_SYMBOL(llog_recov_thread_fini); + +static int llog_recov_thread_replay(struct llog_ctxt *ctxt, + void *cb, void *arg) +{ + struct obd_device *obd = ctxt->loc_obd; + struct llog_process_cat_args *lpca; int rc; ENTRY; - up(&data->llpa_sem); - lock_kernel(); - ptlrpc_daemonize(); /* thread never needs to do IO */ + if (obd->obd_stopping) + RETURN(-ENODEV); - THREAD_NAME(current->comm, sizeof(current->comm) - 1, "llog_process"); - SIGNAL_MASK_LOCK(current, flags); - sigfillset(¤t->blocked); - RECALC_SIGPENDING; - SIGNAL_MASK_UNLOCK(current, flags); - unlock_kernel(); + /* + * This will be balanced in llog_cat_process_thread() + */ + OBD_ALLOC_PTR(lpca); + if (!lpca) + RETURN(-ENOMEM); - rc = llog_open(ctxt, &llh, &logid, NULL, 0); - if (rc) { - CERROR("llog_create failed %d\n", rc); - RETURN(rc); - } - rc = llog_init_handle(llh, LLOG_F_IS_CAT, NULL); - if (rc) { - CERROR("llog_init_handle failed %d\n", rc); - GOTO(out, rc); - } + lpca->lpca_cb = cb; + lpca->lpca_arg = arg; - if (cb) { - rc = llog_cat_process(llh, (llog_cb_t)cb, NULL); - if (rc != LLOG_PROC_BREAK) - CERROR("llog_cat_process failed %d\n", rc); + /* + * This will be balanced in llog_cat_process_thread() + */ + lpca->lpca_ctxt = llog_ctxt_get(ctxt); + if (!lpca->lpca_ctxt) { + OBD_FREE_PTR(lpca); + RETURN(-ENODEV); + } + rc = cfs_create_thread(llog_cat_process_thread, lpca, CFS_DAEMON_FLAGS); + if (rc < 0) { + CERROR("Error starting llog_cat_process_thread(): %d\n", rc); + OBD_FREE_PTR(lpca); + llog_ctxt_put(ctxt); } else { - CWARN("no callback function for recovery\n"); + CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc); + rc = 0; } - CDEBUG(D_HA, "send llcd %p:%p forcibly after recovery\n", - ctxt->loc_llcd, ctxt); - llog_sync(ctxt, NULL); -out: - rc = llog_cat_put(llh); - if (rc) - CERROR("llog_cat_put failed %d\n", rc); - RETURN(rc); } -static int llog_recovery_generic(struct llog_ctxt *ctxt, void *handle,void *arg) +int llog_obd_repl_connect(struct llog_ctxt *ctxt, + struct llog_logid *logid, struct llog_gen *gen, + struct obd_uuid *uuid) { int rc; ENTRY; - down(&llpa.llpa_sem); - llpa.llpa_ctxt = ctxt; - llpa.llpa_cb = handle; - llpa.llpa_arg = arg; + /* + * Send back cached llcd from llog before recovery if we have any. + * This is void is nothing cached is found there. + */ + llog_sync(ctxt, NULL, 0); - rc = kernel_thread(log_process_thread, &llpa, CLONE_VM | CLONE_FILES); - if (rc < 0) - CERROR("error starting log_process_thread: %d\n", rc); - else { - CDEBUG(D_HA, "log_process_thread: %d\n", rc); - rc = 0; - } + /* + * Start recovery in separate thread. + */ + cfs_mutex_lock(&ctxt->loc_mutex); + ctxt->loc_gen = *gen; + rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid); + cfs_mutex_unlock(&ctxt->loc_mutex); RETURN(rc); } +EXPORT_SYMBOL(llog_obd_repl_connect); -int llog_repl_connect(struct llog_ctxt *ctxt, int count, - struct llog_logid *logid, struct llog_gen *gen, - struct obd_uuid *uuid) +/** + * Deleted objects have a commit callback that cancels the MDS + * log record for the deletion. The commit callback calls this + * function. + */ +int llog_obd_repl_cancel(struct llog_ctxt *ctxt, + struct lov_stripe_md *lsm, int count, + struct llog_cookie *cookies, int flags) { + struct llog_commit_master *lcm; struct llog_canceld_ctxt *llcd; - int rc; + int rc = 0; ENTRY; - /* send back llcd before recovery from llog */ - if (ctxt->loc_llcd != NULL) { - CWARN("llcd %p:%p not empty\n", ctxt->loc_llcd, ctxt); - llog_sync(ctxt, NULL); + LASSERT(ctxt != NULL); + + cfs_mutex_lock(&ctxt->loc_mutex); + if (!ctxt->loc_lcm) { + CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt); + GOTO(out, rc = -ENODEV); + } + lcm = ctxt->loc_lcm; + CDEBUG(D_INFO, "cancel on lsm %p\n", lcm); + + /* + * Let's check if we have all structures alive. We also check for + * possible shutdown. Do nothing if we're stopping. + */ + if (ctxt->loc_flags & LLOG_CTXT_FLAG_STOP) { + CDEBUG(D_RPCTRACE, "Last sync was done for ctxt %p\n", ctxt); + GOTO(out, rc = -ENODEV); + } + + if (ctxt->loc_imp == NULL) { + CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt); + GOTO(out, rc = -ENODEV); } - down(&ctxt->loc_sem); - ctxt->loc_gen = *gen; - llcd = llcd_grab(); - if (llcd == NULL) { - CERROR("couldn't get an llcd\n"); - up(&ctxt->loc_sem); - RETURN(-ENOMEM); + if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) { + CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n", + ctxt); + GOTO(out, rc = -ENODEV); } - llcd->llcd_ctxt = ctxt; - ctxt->loc_llcd = llcd; - up(&ctxt->loc_sem); - rc = llog_recovery_generic(ctxt, ctxt->llog_proc_cb, logid); - if (rc != 0) - CERROR("error recovery process: %d\n", rc); + llcd = ctxt->loc_llcd; + + if (count > 0 && cookies != NULL) { + /* + * Get new llcd from ctxt if required. + */ + if (!llcd) { + llcd = llcd_get(ctxt); + if (!llcd) + GOTO(out, rc = -ENOMEM); + /* + * Allocation is successful, let's check for stop + * flag again to fall back as soon as possible. + */ + if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) + GOTO(out, rc = -ENODEV); + } + + /* + * Llcd does not have enough room for @cookies. Let's push + * it out and allocate new one. + */ + if (!llcd_fit(llcd, cookies)) { + rc = llcd_push(ctxt); + if (rc) + GOTO(out, rc); + llcd = llcd_get(ctxt); + if (!llcd) + GOTO(out, rc = -ENOMEM); + /* + * Allocation is successful, let's check for stop + * flag again to fall back as soon as possible. + */ + if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) + GOTO(out, rc = -ENODEV); + } + + /* + * Copy cookies to @llcd, no matter old or new allocated + * one. + */ + llcd_copy(llcd, cookies); + } + + /* + * Let's check if we need to send copied @cookies asap. If yes + * then do it. + */ + if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) { + CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd); + rc = llcd_push(ctxt); + if (rc) + GOTO(out, rc); + } + EXIT; +out: + if (rc) + llcd_put(ctxt); + + if (flags & OBD_LLOG_FL_EXIT) + ctxt->loc_flags = LLOG_CTXT_FLAG_STOP; + + cfs_mutex_unlock(&ctxt->loc_mutex); + return rc; +} +EXPORT_SYMBOL(llog_obd_repl_cancel); + +int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp, + int flags) +{ + int rc = 0; + ENTRY; + /* + * Flush any remaining llcd. + */ + cfs_mutex_lock(&ctxt->loc_mutex); + if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) { + /* + * This is ost->mds connection, we can't be sure that mds + * can still receive cookies, let's killed the cached llcd. + */ + CDEBUG(D_RPCTRACE, "Kill cached llcd\n"); + llcd_put(ctxt); + + if (flags & OBD_LLOG_FL_EXIT) + ctxt->loc_flags = LLOG_CTXT_FLAG_STOP; + + cfs_mutex_unlock(&ctxt->loc_mutex); + } else { + /* + * This is either llog_sync() from generic llog code or sync + * on client disconnect. In either way let's do it and send + * llcds to the target with waiting for completion. + */ + CDEBUG(D_RPCTRACE, "Sync cached llcd\n"); + cfs_mutex_unlock(&ctxt->loc_mutex); + rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW | + flags); + } RETURN(rc); } -EXPORT_SYMBOL(llog_repl_connect); +EXPORT_SYMBOL(llog_obd_repl_sync); #else /* !__KERNEL__ */ -int llog_obd_repl_cancel(struct llog_ctxt *ctxt, int count, - struct llog_cookie *cookies, int flags, void *data) +int llog_obd_repl_cancel(struct llog_ctxt *ctxt, + struct lov_stripe_md *lsm, int count, + struct llog_cookie *cookies, int flags) { return 0; } #endif + +/** + * Module init time fucntion. Initializes slab for llcd objects. + */ +int llog_recov_init(void) +{ + int llcd_size; + + llcd_size = CFS_PAGE_SIZE - + lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL); + llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies); + llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0); + if (!llcd_cache) { + CERROR("Error allocating llcd cache\n"); + return -ENOMEM; + } + return 0; +} + +/** + * Module fini time fucntion. Releases slab for llcd objects. + */ +void llog_recov_fini(void) +{ + /* + * Kill llcd cache when thread is stopped and we're sure no + * llcd in use left. + */ + if (llcd_cache) { + /* + * In 2.6.22 cfs_mem_cache_destroy() will not return error + * for busy resources. Let's check it another way. + */ + LASSERTF(cfs_atomic_read(&llcd_count) == 0, + "Can't destroy llcd cache! Number of " + "busy llcds: %d\n", cfs_atomic_read(&llcd_count)); + cfs_mem_cache_destroy(llcd_cache); + llcd_cache = NULL; + } +}