X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmgc%2Fmgc_request.c;h=e503c1b80594f323788f69069f126cbab93821d3;hp=337418be97857e941e1a2071724ec038318c489f;hb=3e38436dc09097429e1ca1fbfa3ef4981e124715;hpb=56e33bbf23609d3fbc085d60dad18953e0b0c6e7 diff --git a/lustre/mgc/mgc_request.c b/lustre/mgc/mgc_request.c index 337418b..e503c1b 100644 --- a/lustre/mgc/mgc_request.c +++ b/lustre/mgc/mgc_request.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,30 +38,21 @@ * Author: Nathan Rutman */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MGC #define D_MGC D_CONFIG /*|D_WARNING*/ -#ifdef __KERNEL__ -# include -# include -# include -# include -#else -# include -#endif - +#include #include #include #include #include -#include #include +#include + #include "mgc_internal.h" -static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id) +static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id, + int type) { __u64 resname = 0; @@ -75,24 +66,36 @@ static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id) } memcpy(&resname, name, len); - memset(res_id, 0, sizeof(*res_id)); - /* Always use the same endianness for the resid */ + memset(res_id, 0, sizeof(*res_id)); res_id->name[0] = cpu_to_le64(resname); + /* XXX: unfortunately, sptlprc and config llog share one lock */ + switch(type) { + case CONFIG_T_CONFIG: + case CONFIG_T_SPTLRPC: + resname = 0; + break; + case CONFIG_T_RECOVER: + resname = type; + break; + default: + LBUG(); + } + res_id->name[1] = cpu_to_le64(resname); CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name, res_id->name[0], res_id->name[1], (char *)&res_id->name[0]); return 0; } -int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id) +int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type) { /* fsname is at most 8 chars long, maybe contain "-". * e.g. "lustre", "SUN-000" */ - return mgc_name2resid(fsname, strlen(fsname), res_id); + return mgc_name2resid(fsname, strlen(fsname), res_id, type); } EXPORT_SYMBOL(mgc_fsname2resid); -int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id) +int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type) { char *name_end; int len; @@ -102,22 +105,20 @@ int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id) name_end = strrchr(logname, '-'); LASSERT(name_end); len = name_end - logname; - return mgc_name2resid(logname, len, res_id); + return mgc_name2resid(logname, len, res_id, type); } /********************** config llog list **********************/ static CFS_LIST_HEAD(config_llog_list); -static spinlock_t config_list_lock = SPIN_LOCK_UNLOCKED; +static DEFINE_SPINLOCK(config_list_lock); /* Take a reference to a config log */ static int config_log_get(struct config_llog_data *cld) { ENTRY; - if (cld->cld_stopping) - RETURN(1); - atomic_inc(&cld->cld_refcount); + cfs_atomic_inc(&cld->cld_refcount); CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname, - atomic_read(&cld->cld_refcount)); + cfs_atomic_read(&cld->cld_refcount)); RETURN(0); } @@ -128,30 +129,25 @@ static void config_log_put(struct config_llog_data *cld) ENTRY; CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname, - atomic_read(&cld->cld_refcount)); - LASSERT(atomic_read(&cld->cld_refcount) > 0); + cfs_atomic_read(&cld->cld_refcount)); + LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0); /* spinlock to make sure no item with 0 refcount in the list */ - spin_lock(&config_list_lock); - if (unlikely(atomic_dec_and_test(&cld->cld_refcount))) { - list_del(&cld->cld_list_chain); - spin_unlock(&config_list_lock); + if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) { + cfs_list_del(&cld->cld_list_chain); + spin_unlock(&config_list_lock); CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname); + if (cld->cld_recover) + config_log_put(cld->cld_recover); if (cld->cld_sptlrpc) config_log_put(cld->cld_sptlrpc); - if (cld->cld_is_sptlrpc) + if (cld_is_sptlrpc(cld)) sptlrpc_conf_log_stop(cld->cld_logname); class_export_put(cld->cld_mgcexp); - OBD_FREE(cld->cld_logname, strlen(cld->cld_logname) + 1); - if (cld->cld_cfg.cfg_instance != NULL) - OBD_FREE(cld->cld_cfg.cfg_instance, - strlen(cld->cld_cfg.cfg_instance) + 1); - OBD_FREE(cld, sizeof(*cld)); - } else { - spin_unlock(&config_list_lock); + OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1); } EXIT; @@ -163,43 +159,37 @@ struct config_llog_data *config_log_find(char *logname, struct config_llog_instance *cfg) { struct config_llog_data *cld; - char *logid = logname; - int match_instance = 0; + struct config_llog_data *found = NULL; + void * instance; ENTRY; - if (cfg && cfg->cfg_instance) { - match_instance++; - logid = cfg->cfg_instance; - } - if (!logid) { - CERROR("No log specified\n"); - RETURN(ERR_PTR(-EINVAL)); - } + LASSERT(logname != NULL); + + instance = cfg ? cfg->cfg_instance : NULL; + spin_lock(&config_list_lock); + cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) { + /* check if instance equals */ + if (instance != cld->cld_cfg.cfg_instance) + continue; - spin_lock(&config_list_lock); - list_for_each_entry(cld, &config_llog_list, cld_list_chain) { - if (match_instance && cld->cld_cfg.cfg_instance && - strcmp(logid, cld->cld_cfg.cfg_instance) == 0) - goto out_found; - if (!match_instance && - strcmp(logid, cld->cld_logname) == 0) - goto out_found; + /* instance may be NULL, should check name */ + if (strcmp(logname, cld->cld_logname) == 0) { + found = cld; + break; + } } - spin_unlock(&config_list_lock); - - CDEBUG(D_CONFIG, "can't get log %s\n", logid); - RETURN(ERR_PTR(-ENOENT)); -out_found: - atomic_inc(&cld->cld_refcount); - spin_unlock(&config_list_lock); - LASSERT(cld->cld_stopping == 0 || cld->cld_is_sptlrpc == 0); - RETURN(cld); + if (found) { + cfs_atomic_inc(&found->cld_refcount); + LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0); + } + spin_unlock(&config_list_lock); + RETURN(found); } static struct config_llog_data *do_config_log_add(struct obd_device *obd, char *logname, - unsigned int is_sptlrpc, + int type, struct config_llog_instance *cfg, struct super_block *sb) { @@ -207,62 +197,90 @@ struct config_llog_data *do_config_log_add(struct obd_device *obd, int rc; ENTRY; - CDEBUG(D_MGC, "do adding config log %s:%s\n", logname, - cfg ? cfg->cfg_instance : "NULL"); + CDEBUG(D_MGC, "do adding config log %s:%p\n", logname, + cfg ? cfg->cfg_instance : 0); - OBD_ALLOC(cld, sizeof(*cld)); + OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1); if (!cld) RETURN(ERR_PTR(-ENOMEM)); - OBD_ALLOC(cld->cld_logname, strlen(logname) + 1); - if (!cld->cld_logname) { - OBD_FREE(cld, sizeof(*cld)); - RETURN(ERR_PTR(-ENOMEM)); - } + strcpy(cld->cld_logname, logname); if (cfg) cld->cld_cfg = *cfg; + else + cld->cld_cfg.cfg_callback = class_config_llog_handler; + mutex_init(&cld->cld_lock); cld->cld_cfg.cfg_last_idx = 0; cld->cld_cfg.cfg_flags = 0; cld->cld_cfg.cfg_sb = sb; - cld->cld_is_sptlrpc = is_sptlrpc; - atomic_set(&cld->cld_refcount, 1); + cld->cld_type = type; + cfs_atomic_set(&cld->cld_refcount, 1); /* Keep the mgc around until we are done */ cld->cld_mgcexp = class_export_get(obd->obd_self_export); - if (cfg && cfg->cfg_instance != NULL) { - OBD_ALLOC(cld->cld_cfg.cfg_instance, - strlen(cfg->cfg_instance) + 1); - strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance); - } - - if (is_sptlrpc) { + if (cld_is_sptlrpc(cld)) { sptlrpc_conf_log_start(logname); cld->cld_cfg.cfg_obdname = obd->obd_name; } - rc = mgc_logname2resid(logname, &cld->cld_resid); + rc = mgc_logname2resid(logname, &cld->cld_resid, type); - spin_lock(&config_list_lock); - list_add(&cld->cld_list_chain, &config_llog_list); - spin_unlock(&config_list_lock); + spin_lock(&config_list_lock); + cfs_list_add(&cld->cld_list_chain, &config_llog_list); + spin_unlock(&config_list_lock); if (rc) { config_log_put(cld); RETURN(ERR_PTR(rc)); } - if (is_sptlrpc) { + if (cld_is_sptlrpc(cld)) { rc = mgc_process_log(obd, cld); - if (rc) + if (rc && rc != -ENOENT) CERROR("failed processing sptlrpc log: %d\n", rc); } RETURN(cld); } -/** - * Add this log to our list of active logs. +static struct config_llog_data *config_recover_log_add(struct obd_device *obd, + char *fsname, + struct config_llog_instance *cfg, + struct super_block *sb) +{ + struct config_llog_instance lcfg = *cfg; + struct lustre_sb_info *lsi = s2lsi(sb); + struct config_llog_data *cld; + char logname[32]; + + if (IS_OST(lsi)) + return NULL; + + /* for osp-on-ost, see lustre_start_osp() */ + if (IS_MDT(lsi) && lcfg.cfg_instance) + return NULL; + + /* we have to use different llog for clients and mdts for cmd + * where only clients are notified if one of cmd server restarts */ + LASSERT(strlen(fsname) < sizeof(logname) / 2); + strcpy(logname, fsname); + if (IS_SERVER(lsi)) { /* mdt */ + LASSERT(lcfg.cfg_instance == NULL); + lcfg.cfg_instance = sb; + strcat(logname, "-mdtir"); + } else { + LASSERT(lcfg.cfg_instance != NULL); + strcat(logname, "-cliir"); + } + + cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb); + return cld; +} + + +/** Add this log to the list of active logs watched by an MGC. + * Active means we're watching for updates. * We have one active log per "mount" - client instance or servername. * Each instance may be at a different point in the log. */ @@ -270,12 +288,14 @@ static int config_log_add(struct obd_device *obd, char *logname, struct config_llog_instance *cfg, struct super_block *sb) { - struct config_llog_data *cld, *sptlrpc_cld; - char seclogname[20]; + struct lustre_sb_info *lsi = s2lsi(sb); + struct config_llog_data *cld; + struct config_llog_data *sptlrpc_cld; + char seclogname[32]; char *ptr; ENTRY; - CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance); + CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance); /* * for each regular log, the depended sptlrpc log name is @@ -291,15 +311,16 @@ static int config_log_add(struct obd_device *obd, char *logname, strcpy(seclogname + (ptr - logname), "-sptlrpc"); sptlrpc_cld = config_log_find(seclogname, NULL); - if (IS_ERR(sptlrpc_cld)) { - sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL); + if (sptlrpc_cld == NULL) { + sptlrpc_cld = do_config_log_add(obd, seclogname, + CONFIG_T_SPTLRPC, NULL, NULL); if (IS_ERR(sptlrpc_cld)) { CERROR("can't create sptlrpc log: %s\n", seclogname); RETURN(PTR_ERR(sptlrpc_cld)); } } - cld = do_config_log_add(obd, logname, 0, cfg, sb); + cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb); if (IS_ERR(cld)) { CERROR("can't create log: %s\n", logname); config_log_put(sptlrpc_cld); @@ -308,23 +329,38 @@ static int config_log_add(struct obd_device *obd, char *logname, cld->cld_sptlrpc = sptlrpc_cld; + LASSERT(lsi->lsi_lmd); + if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) { + struct config_llog_data *recover_cld; + *strrchr(seclogname, '-') = 0; + recover_cld = config_recover_log_add(obd, seclogname, cfg, sb); + if (IS_ERR(recover_cld)) { + config_log_put(cld); + RETURN(PTR_ERR(recover_cld)); + } + cld->cld_recover = recover_cld; + } + RETURN(0); } -DECLARE_MUTEX(llog_process_lock); +DEFINE_MUTEX(llog_process_lock); -/* Stop watching for updates on this log. */ +/** Stop watching for updates on this log. + */ static int config_log_end(char *logname, struct config_llog_instance *cfg) { - struct config_llog_data *cld, *cld_sptlrpc = NULL; + struct config_llog_data *cld; + struct config_llog_data *cld_sptlrpc = NULL; + struct config_llog_data *cld_recover = NULL; int rc = 0; ENTRY; cld = config_log_find(logname, cfg); - if (IS_ERR(cld)) - RETURN(PTR_ERR(cld)); + if (cld == NULL) + RETURN(-ENOENT); - down(&llog_process_lock); + mutex_lock(&cld->cld_lock); /* * if cld_stopping is set, it means we didn't start the log thus * not owning the start ref. this can happen after previous umount: @@ -333,19 +369,29 @@ static int config_log_end(char *logname, struct config_llog_instance *cfg) * calling start_log. */ if (unlikely(cld->cld_stopping)) { - up(&llog_process_lock); + mutex_unlock(&cld->cld_lock); /* drop the ref from the find */ config_log_put(cld); RETURN(rc); } cld->cld_stopping = 1; - up(&llog_process_lock); - spin_lock(&config_list_lock); - cld_sptlrpc = cld->cld_sptlrpc; - cld->cld_sptlrpc = NULL; - spin_unlock(&config_list_lock); + cld_recover = cld->cld_recover; + cld->cld_recover = NULL; + mutex_unlock(&cld->cld_lock); + + if (cld_recover) { + mutex_lock(&cld_recover->cld_lock); + cld_recover->cld_stopping = 1; + mutex_unlock(&cld_recover->cld_lock); + config_log_put(cld_recover); + } + + spin_lock(&config_list_lock); + cld_sptlrpc = cld->cld_sptlrpc; + cld->cld_sptlrpc = NULL; + spin_unlock(&config_list_lock); if (cld_sptlrpc) config_log_put(cld_sptlrpc); @@ -360,20 +406,52 @@ static int config_log_end(char *logname, struct config_llog_instance *cfg) RETURN(rc); } +int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off, + int count, int *eof, void *data) +{ + struct obd_device *obd = data; + struct obd_import *imp = obd->u.cli.cl_import; + struct obd_connect_data *ocd = &imp->imp_connect_data; + struct config_llog_data *cld; + int rc = 0; + ENTRY; + + rc = snprintf(page, count, "imperative_recovery: %s\n", + OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED"); + rc += snprintf(page + rc, count - rc, "client_state:\n"); + + spin_lock(&config_list_lock); + cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) { + if (cld->cld_recover == NULL) + continue; + rc += snprintf(page + rc, count - rc, + " - { client: %s, nidtbl_version: %u }\n", + cld->cld_logname, + cld->cld_recover->cld_cfg.cfg_last_idx); + } + spin_unlock(&config_list_lock); + + RETURN(rc); +} + /* reenqueue any lost locks */ #define RQ_RUNNING 0x1 #define RQ_NOW 0x2 #define RQ_LATER 0x4 #define RQ_STOP 0x8 -static int rq_state = 0; -static cfs_waitq_t rq_waitq; - -static int mgc_requeue_add(struct config_llog_data *cld, int later); +static int rq_state = 0; +static cfs_waitq_t rq_waitq; +static DECLARE_COMPLETION(rq_exit); static void do_requeue(struct config_llog_data *cld) { - LASSERT(atomic_read(&cld->cld_refcount) > 0); + ENTRY; + LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0); + /* Do not run mgc_process_log on a disconnected export or an + export which is being disconnected. Take the client + semaphore to make the check non-racy. */ + down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem); if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) { CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname); mgc_process_log(cld->cld_mgcexp->exp_obd, cld); @@ -381,40 +459,46 @@ static void do_requeue(struct config_llog_data *cld) CDEBUG(D_MGC, "disconnecting, won't update log %s\n", cld->cld_logname); } + up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem); - /* Whether we enqueued again or not in mgc_process_log, we're done - * with the ref from the old enqueue */ - config_log_put(cld); + EXIT; } +/* this timeout represents how many seconds MGC should wait before + * requeue config and recover lock to the MGS. We need to randomize this + * in order to not flood the MGS. + */ +#define MGC_TIMEOUT_MIN_SECONDS 5 +#define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */ + static int mgc_requeue_thread(void *data) { - struct l_wait_info lwi_now, lwi_later; - struct config_llog_data *cld, *cld_next, *cld_prev; - char name[] = "ll_cfg_requeue"; int rc = 0; ENTRY; - cfs_daemonize(name); - CDEBUG(D_MGC, "Starting requeue thread\n"); - lwi_later = LWI_TIMEOUT(60 * HZ, NULL, NULL); - l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later); - /* Keep trying failed locks periodically */ - spin_lock(&config_list_lock); - while (rq_state & (RQ_NOW | RQ_LATER)) { - /* Any new or requeued lostlocks will change the state */ - rq_state &= ~(RQ_NOW | RQ_LATER); - spin_unlock(&config_list_lock); + spin_lock(&config_list_lock); + rq_state |= RQ_RUNNING; + while (1) { + struct l_wait_info lwi; + struct config_llog_data *cld, *cld_prev; + int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC; + int stopped = !!(rq_state & RQ_STOP); + int to; + + /* Any new or requeued lostlocks will change the state */ + rq_state &= ~(RQ_NOW | RQ_LATER); + spin_unlock(&config_list_lock); /* Always wait a few seconds to allow the server who caused the lock revocation to finish its setup, plus some random so everyone doesn't try to reconnect at once. */ - lwi_now = LWI_TIMEOUT(3 * HZ + (ll_rand() & 0xff) * (HZ / 100), - NULL, NULL); - l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now); + to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ; + to += rand * CFS_HZ / 100; /* rand is centi-seconds */ + lwi = LWI_TIMEOUT(to, NULL, NULL); + l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi); /* * iterate & processing through the list. for each cld, process @@ -423,203 +507,280 @@ static int mgc_requeue_thread(void *data) * it's guaranteed any item in the list must have * reference > 0; and if cld_lostlock is set, at * least one reference is taken by the previous enqueue. - * - * Note: releasing a cld might lead to itself and its depended - * sptlrpc cld be unlinked from the list. to safely iterate - * we need to take a reference on next cld before processing. */ cld_prev = NULL; - spin_lock(&config_list_lock); - list_for_each_entry_safe(cld, cld_next, &config_llog_list, - cld_list_chain) { - if (cld->cld_list_chain.next != &config_llog_list) - atomic_inc(&cld_next->cld_refcount); - - if (cld->cld_lostlock) { - if (cld->cld_sptlrpc && - cld->cld_sptlrpc->cld_lostlock) { - cld->cld_sptlrpc->cld_lostlock = 0; - - spin_unlock(&config_list_lock); - do_requeue(cld->cld_sptlrpc); - spin_lock(&config_list_lock); - LASSERT(cld->cld_lostlock); - } + spin_lock(&config_list_lock); + cfs_list_for_each_entry(cld, &config_llog_list, + cld_list_chain) { + if (!cld->cld_lostlock) + continue; - cld->cld_lostlock = 0; - - spin_unlock(&config_list_lock); - do_requeue(cld); - spin_lock(&config_list_lock); - } + spin_unlock(&config_list_lock); + LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0); - if (cld_prev) { - spin_unlock(&config_list_lock); + /* Whether we enqueued again or not in mgc_process_log, + * we're done with the ref from the old enqueue */ + if (cld_prev) config_log_put(cld_prev); - spin_lock(&config_list_lock); - } + cld_prev = cld; - cld_prev = cld_next; - } - spin_unlock(&config_list_lock); - - /* Wait a bit to see if anyone else needs a requeue */ - l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), - &lwi_later); - spin_lock(&config_list_lock); - } - /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */ - rq_state &= ~RQ_RUNNING; - spin_unlock(&config_list_lock); + cld->cld_lostlock = 0; + if (likely(!stopped)) + do_requeue(cld); - CDEBUG(D_MGC, "Ending requeue thread\n"); - RETURN(rc); + spin_lock(&config_list_lock); + } + spin_unlock(&config_list_lock); + if (cld_prev) + config_log_put(cld_prev); + + /* break after scanning the list so that we can drop + * refcount to losing lock clds */ + if (unlikely(stopped)) { + spin_lock(&config_list_lock); + break; + } + + /* Wait a bit to see if anyone else needs a requeue */ + lwi = (struct l_wait_info) { 0 }; + l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), + &lwi); + spin_lock(&config_list_lock); + } + /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */ + rq_state &= ~RQ_RUNNING; + spin_unlock(&config_list_lock); + + complete(&rq_exit); + + CDEBUG(D_MGC, "Ending requeue thread\n"); + RETURN(rc); } /* Add a cld to the list to requeue. Start the requeue thread if needed. We are responsible for dropping the config log reference from here on out. */ -static int mgc_requeue_add(struct config_llog_data *cld, int later) +static void mgc_requeue_add(struct config_llog_data *cld) { - int rc = 0; + ENTRY; - CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n", - cld->cld_logname, later, atomic_read(&cld->cld_refcount), + CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n", + cld->cld_logname, cfs_atomic_read(&cld->cld_refcount), cld->cld_stopping, rq_state); - LASSERT(atomic_read(&cld->cld_refcount) > 0); + LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0); + + mutex_lock(&cld->cld_lock); + if (cld->cld_stopping || cld->cld_lostlock) { + mutex_unlock(&cld->cld_lock); + RETURN_EXIT; + } + /* this refcount will be released in mgc_requeue_thread. */ + config_log_get(cld); + cld->cld_lostlock = 1; + mutex_unlock(&cld->cld_lock); + + /* Hold lock for rq_state */ + spin_lock(&config_list_lock); + if (rq_state & RQ_STOP) { + spin_unlock(&config_list_lock); + cld->cld_lostlock = 0; + config_log_put(cld); + } else { + rq_state |= RQ_NOW; + spin_unlock(&config_list_lock); + cfs_waitq_signal(&rq_waitq); + } + EXIT; +} - /* Hold lock for rq_state */ - spin_lock(&config_list_lock); +/********************** class fns **********************/ +static int mgc_local_llog_init(const struct lu_env *env, + struct obd_device *obd, + struct obd_device *disk) +{ + struct llog_ctxt *ctxt; + int rc; - if (cld->cld_stopping || (rq_state & RQ_STOP)) { - cld->cld_lostlock = 0; - spin_unlock(&config_list_lock); - config_log_put(cld); - RETURN(0); - } + ENTRY; - cld->cld_lostlock = 1; + rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk, + &llog_osd_ops); + if (rc) + RETURN(rc); - if (!(rq_state & RQ_RUNNING)) { - LASSERT(rq_state == 0); - rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW); - spin_unlock(&config_list_lock); - rc = cfs_kernel_thread(mgc_requeue_thread, 0, - CLONE_VM | CLONE_FILES); - if (rc < 0) { - CERROR("log %s: cannot start requeue thread (%d)," - "no more log updates!\n", cld->cld_logname, rc); - /* Drop the ref, since the rq thread won't */ - cld->cld_lostlock = 0; - config_log_put(cld); - rq_state = 0; - RETURN(rc); - } - } else { - rq_state |= later ? RQ_LATER : RQ_NOW; - spin_unlock(&config_list_lock); - cfs_waitq_signal(&rq_waitq); - } + ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT); + LASSERT(ctxt); + ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir; + llog_ctxt_put(ctxt); - RETURN(0); + RETURN(0); } -/********************** class fns **********************/ +static int mgc_local_llog_fini(const struct lu_env *env, + struct obd_device *obd) +{ + struct llog_ctxt *ctxt; -static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb, - struct vfsmount *mnt) + ENTRY; + + ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT); + llog_cleanup(env, ctxt); + + RETURN(0); +} + +static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb) { - struct lvfs_run_ctxt saved; - struct lustre_sb_info *lsi = s2lsi(sb); - struct client_obd *cli = &obd->u.cli; - struct dentry *dentry; - char *label; - int err = 0; - ENTRY; + struct lustre_sb_info *lsi = s2lsi(sb); + struct client_obd *cli = &obd->u.cli; + struct lu_fid rfid, fid; + struct dt_object *root, *dto; + struct lu_env *env; + int rc = 0; + + ENTRY; + + LASSERT(lsi); + LASSERT(lsi->lsi_dt_dev); + + OBD_ALLOC_PTR(env); + if (env == NULL) + RETURN(-ENOMEM); + + /* The mgc fs exclusion sem. Only one fs can be setup at a time. */ + down(&cli->cl_mgc_sem); + + cfs_cleanup_group_info(); + + /* Setup the configs dir */ + rc = lu_env_init(env, LCT_MG_THREAD); + if (rc) + GOTO(out_err, rc); + + fid.f_seq = FID_SEQ_LOCAL_NAME; + fid.f_oid = 1; + fid.f_ver = 0; + rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid, + &cli->cl_mgc_los); + if (rc) + GOTO(out_env, rc); + + rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid); + if (rc) + GOTO(out_env, rc); + + root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid, + &cli->cl_mgc_los->los_dev->dd_lu_dev); + if (unlikely(IS_ERR(root))) + GOTO(out_los, rc = PTR_ERR(root)); + + dto = local_file_find_or_create(env, cli->cl_mgc_los, root, + MOUNT_CONFIGS_DIR, + S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO); + lu_object_put_nocache(env, &root->do_lu); + if (IS_ERR(dto)) + GOTO(out_los, rc = PTR_ERR(dto)); + + cli->cl_mgc_configs_dir = dto; + + LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt); + rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd); + if (rc) + GOTO(out_llog, rc); + + /* We take an obd ref to insure that we can't get to mgc_cleanup + * without calling mgc_fs_cleanup first. */ + class_incref(obd, "mgc_fs", obd); + + /* We keep the cl_mgc_sem until mgc_fs_cleanup */ + EXIT; +out_llog: + if (rc) { + lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu); + cli->cl_mgc_configs_dir = NULL; + } +out_los: + if (rc < 0) { + local_oid_storage_fini(env, cli->cl_mgc_los); + cli->cl_mgc_los = NULL; + up(&cli->cl_mgc_sem); + } +out_env: + lu_env_fini(env); +out_err: + OBD_FREE_PTR(env); + return rc; +} - LASSERT(lsi); - LASSERT(lsi->lsi_srv_mnt == mnt); +static int mgc_fs_cleanup(struct obd_device *obd) +{ + struct lu_env env; + struct client_obd *cli = &obd->u.cli; + int rc; - /* The mgc fs exclusion sem. Only one fs can be setup at a time. */ - down(&cli->cl_mgc_sem); + ENTRY; - cleanup_group_info(); + LASSERT(cli->cl_mgc_los != NULL); - obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd)); - if (IS_ERR(obd->obd_fsops)) { - up(&cli->cl_mgc_sem); - CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd), - PTR_ERR(obd->obd_fsops)); - RETURN(PTR_ERR(obd->obd_fsops)); - } + rc = lu_env_init(&env, LCT_MG_THREAD); + if (rc) + GOTO(unlock, rc); - cli->cl_mgc_vfsmnt = mnt; - fsfilt_setup(obd, mnt->mnt_sb); - - OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt); - obd->obd_lvfs_ctxt.pwdmnt = mnt; - obd->obd_lvfs_ctxt.pwd = mnt->mnt_root; - obd->obd_lvfs_ctxt.fs = get_ds(); - - push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - dentry = lookup_one_len(MOUNT_CONFIGS_DIR, current->fs->pwd, - strlen(MOUNT_CONFIGS_DIR)); - pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - if (IS_ERR(dentry)) { - err = PTR_ERR(dentry); - CERROR("cannot lookup %s directory: rc = %d\n", - MOUNT_CONFIGS_DIR, err); - GOTO(err_ops, err); - } - cli->cl_mgc_configs_dir = dentry; + mgc_local_llog_fini(&env, obd); - /* We take an obd ref to insure that we can't get to mgc_cleanup - without calling mgc_fs_cleanup first. */ - class_incref(obd, "mgc_fs", obd); + lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu); + cli->cl_mgc_configs_dir = NULL; - label = fsfilt_get_label(obd, mnt->mnt_sb); - if (label) - CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label); + local_oid_storage_fini(&env, cli->cl_mgc_los); + cli->cl_mgc_los = NULL; + lu_env_fini(&env); - /* We keep the cl_mgc_sem until mgc_fs_cleanup */ - RETURN(0); +unlock: + class_decref(obd, "mgc_fs", obd); + up(&cli->cl_mgc_sem); -err_ops: - fsfilt_put_ops(obd->obd_fsops); - obd->obd_fsops = NULL; - cli->cl_mgc_vfsmnt = NULL; - up(&cli->cl_mgc_sem); - RETURN(err); + RETURN(0); } -static int mgc_fs_cleanup(struct obd_device *obd) +static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd) { - struct client_obd *cli = &obd->u.cli; - int rc = 0; - ENTRY; + struct llog_ctxt *ctxt; + int rc; - LASSERT(cli->cl_mgc_vfsmnt != NULL); + ENTRY; - if (cli->cl_mgc_configs_dir != NULL) { - struct lvfs_run_ctxt saved; - push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - l_dput(cli->cl_mgc_configs_dir); - cli->cl_mgc_configs_dir = NULL; - pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - class_decref(obd, "mgc_fs", obd); - } + /* setup only remote ctxt, the local disk context is switched per each + * filesystem during mgc_fs_setup() */ + rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd, + &llog_client_ops); + if (rc) + RETURN(rc); - cli->cl_mgc_vfsmnt = NULL; - if (obd->obd_fsops) - fsfilt_put_ops(obd->obd_fsops); + ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); + LASSERT(ctxt); - up(&cli->cl_mgc_sem); + llog_initiator_connect(ctxt); + llog_ctxt_put(ctxt); - RETURN(rc); + RETURN(0); +} + +static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd) +{ + struct llog_ctxt *ctxt; + + ENTRY; + + ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); + if (ctxt) + llog_cleanup(env, ctxt); + + RETURN(0); } -static atomic_t mgc_count = ATOMIC_INIT(0); + +static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0); static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage) { int rc = 0; @@ -629,15 +790,21 @@ static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage) case OBD_CLEANUP_EARLY: break; case OBD_CLEANUP_EXPORTS: - if (atomic_dec_and_test(&mgc_count)) { - /* Kick the requeue waitq - cld's should all be - stopping */ - spin_lock(&config_list_lock); - rq_state |= RQ_STOP; - spin_unlock(&config_list_lock); - cfs_waitq_signal(&rq_waitq); + if (cfs_atomic_dec_and_test(&mgc_count)) { + int running; + /* stop requeue thread */ + spin_lock(&config_list_lock); + running = rq_state & RQ_RUNNING; + if (running) + rq_state |= RQ_STOP; + spin_unlock(&config_list_lock); + if (running) { + cfs_waitq_signal(&rq_waitq); + wait_for_completion(&rq_exit); + } } - rc = obd_llog_finish(obd, 0); + obd_cleanup_client_import(obd); + rc = mgc_llog_fini(NULL, obd); if (rc != 0) CERROR("failed to cleanup llogging subsystems\n"); break; @@ -647,12 +814,9 @@ static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage) static int mgc_cleanup(struct obd_device *obd) { - struct client_obd *cli = &obd->u.cli; int rc; ENTRY; - LASSERT(cli->cl_mgc_vfsmnt == NULL); - /* COMPAT_146 - old config logs may have added profiles we don't know about */ if (obd->obd_type->typ_refcnt <= 1) @@ -678,7 +842,7 @@ static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg) if (rc) GOTO(err_decref, rc); - rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL); + rc = mgc_llog_init(NULL, obd); if (rc) { CERROR("failed to setup llogging subsystems\n"); GOTO(err_cleanup, rc); @@ -688,13 +852,22 @@ static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg) lprocfs_obd_setup(obd, lvars.obd_vars); sptlrpc_lprocfs_cliobd_attach(obd); - spin_lock(&config_list_lock); - atomic_inc(&mgc_count); - if (atomic_read(&mgc_count) == 1) { - rq_state &= ~RQ_STOP; - cfs_waitq_init(&rq_waitq); + if (cfs_atomic_inc_return(&mgc_count) == 1) { + rq_state = 0; + cfs_waitq_init(&rq_waitq); + + /* start requeue thread */ + rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL, + "ll_cfg_requeue")); + if (IS_ERR_VALUE(rc)) { + CERROR("%s: Cannot start requeue thread (%d)," + "no more log updates!\n", + obd->obd_name, rc); + GOTO(err_cleanup, rc); + } + /* rc is the task_struct pointer of mgc_requeue_thread. */ + rc = 0; } - spin_unlock(&config_list_lock); RETURN(rc); @@ -719,20 +892,23 @@ static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, /* mgs wants the lock, give it up... */ LDLM_DEBUG(lock, "MGC blocking CB"); ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); - break; - case LDLM_CB_CANCELING: { - /* We've given up the lock, prepare ourselves to update. */ - LDLM_DEBUG(lock, "MGC cancel CB"); + rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); + break; + case LDLM_CB_CANCELING: + /* We've given up the lock, prepare ourselves to update. */ + LDLM_DEBUG(lock, "MGC cancel CB"); - CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n", - lock->l_resource->lr_name.name[0], - (char *)&lock->l_resource->lr_name.name[0]); + CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n", + PLDLMRES(lock->l_resource), + (char *)&lock->l_resource->lr_name.name[0]); if (!cld) { - CERROR("missing data, won't requeue\n"); + CDEBUG(D_INFO, "missing data, won't requeue\n"); break; } + + /* held at mgc_process_log(). */ + LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0); /* Are we done with this log? */ if (cld->cld_stopping) { CDEBUG(D_MGC, "log %s: stopping, won't requeue\n", @@ -744,66 +920,57 @@ static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, (we get called from client_disconnect_export) */ if (!lock->l_conn_export || !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) { - CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n", + CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n", cld->cld_logname); config_log_put(cld); break; } - /* Did we fail to get the lock? */ - if (lock->l_req_mode != lock->l_granted_mode && - !cld->cld_is_sptlrpc) { - CDEBUG(D_MGC, "log %s: original grant failed, will " - "requeue later\n", cld->cld_logname); - /* Try to re-enqueue later */ - rc = mgc_requeue_add(cld, 1); - break; - } + /* Re-enqueue now */ - rc = mgc_requeue_add(cld, 0); + mgc_requeue_add(cld); + config_log_put(cld); break; - } default: LBUG(); } - - if (rc) { - CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ? - "blocking" : "cancel", rc); - LDLM_ERROR(lock, "MGC ast"); - } RETURN(rc); } +/* Not sure where this should go... */ +#define MGC_ENQUEUE_LIMIT 50 +#define MGC_TARGET_REG_LIMIT 10 +#define MGC_SEND_PARAM_LIMIT 10 + /* Send parameter to MGS*/ static int mgc_set_mgs_param(struct obd_export *exp, struct mgs_send_param *msp) { struct ptlrpc_request *req; struct mgs_send_param *req_msp, *rep_msp; - int size[] = { sizeof(struct ptlrpc_body), sizeof(*req_msp) }; - __u32 rep_size[] = { sizeof(struct ptlrpc_body), sizeof(*msp) }; int rc; ENTRY; - req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_MGS_VERSION, - MGS_SET_INFO, 2, size, NULL); + req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), + &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION, + MGS_SET_INFO); if (!req) RETURN(-ENOMEM); - req_msp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*req_msp)); + req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM); if (!req_msp) { ptlrpc_req_finished(req); RETURN(-ENOMEM); } memcpy(req_msp, msp, sizeof(*req_msp)); - ptlrpc_req_set_repsize(req, 2, rep_size); + ptlrpc_request_set_replen(req); + /* Limit how long we will wait for the enqueue to complete */ + req->rq_delay_limit = MGC_SEND_PARAM_LIMIT; rc = ptlrpc_queue_wait(req); if (!rc) { - rep_msp = lustre_swab_repbuf(req, REPLY_REC_OFF, - sizeof(*rep_msp), NULL); + rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM); memcpy(msp, rep_msp, sizeof(*rep_msp)); } @@ -815,32 +982,49 @@ static int mgc_set_mgs_param(struct obd_export *exp, /* Take a config lock so we can get cancel notifications */ static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm, __u32 type, ldlm_policy_data_t *policy, __u32 mode, - int *flags, void *bl_cb, void *cp_cb, void *gl_cb, - void *data, __u32 lvb_len, void *lvb_swabber, - struct lustre_handle *lockh) + __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb, + void *data, __u32 lvb_len, void *lvb_swabber, + struct lustre_handle *lockh) { - struct config_llog_data *cld = (struct config_llog_data *)data; - struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast, - ldlm_completion_ast, NULL, NULL, data}; - - int rc; - ENTRY; + struct config_llog_data *cld = (struct config_llog_data *)data; + struct ldlm_enqueue_info einfo = { + .ei_type = type, + .ei_mode = mode, + .ei_cb_bl = mgc_blocking_ast, + .ei_cb_cp = ldlm_completion_ast, + }; + struct ptlrpc_request *req; + int short_limit = cld_is_sptlrpc(cld); + int rc; + ENTRY; CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname, cld->cld_resid.name[0]); - /* We can only drop this config log ref when we drop the lock */ - if (config_log_get(cld)) - RETURN(ELDLM_LOCK_ABORTED); - /* We need a callback for every lockholder, so don't try to ldlm_lock_match (see rev 1.1.2.11.2.47) */ + req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), + &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION, + LDLM_ENQUEUE); + if (req == NULL) + RETURN(-ENOMEM); + + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0); + ptlrpc_request_set_replen(req); - rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid, - NULL, flags, NULL, 0, NULL, lockh, 0); + /* check if this is server or client */ + if (cld->cld_cfg.cfg_sb) { + struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb); + if (lsi && IS_SERVER(lsi)) + short_limit = 1; + } + /* Limit how long we will wait for the enqueue to complete */ + req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT; + rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags, + NULL, 0, LVB_T_NONE, lockh, 0); /* A failed enqueue should still call the mgc_blocking_ast, where it will be requeued if needed ("grant failed"). */ - + ptlrpc_req_finished(req); RETURN(rc); } @@ -854,59 +1038,16 @@ static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md, RETURN(0); } -#if 0 -static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len, - void *karg, void *uarg) +static void mgc_notify_active(struct obd_device *unused) { - struct obd_device *obd = exp->exp_obd; - struct obd_ioctl_data *data = karg; - struct llog_ctxt *ctxt; - struct lvfs_run_ctxt saved; - int rc; - ENTRY; - - if (!try_module_get(THIS_MODULE)) { - CERROR("Can't get module. Is it alive?"); - return -EINVAL; - } - switch (cmd) { - /* REPLicator context */ - case OBD_IOC_PARSE: { - CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1); - ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT); - rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL); - GOTO(out, rc); - } -#ifdef __KERNEL__ - case OBD_IOC_LLOG_INFO: - case OBD_IOC_LLOG_PRINT: { - ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); - rc = llog_ioctl(ctxt, cmd, data); - - GOTO(out, rc); - } -#endif - /* ORIGinator context */ - case OBD_IOC_DUMP_LOG: { - ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT); - push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL); - pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - if (rc) - RETURN(rc); + /* wakeup mgc_requeue_thread to requeue mgc lock */ + spin_lock(&config_list_lock); + rq_state |= RQ_NOW; + spin_unlock(&config_list_lock); + cfs_waitq_signal(&rq_waitq); - GOTO(out, rc); - } - default: - CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd); - GOTO(out, rc = -ENOTTY); - } -out: - module_put(THIS_MODULE); - - return rc; + /* TODO: Help the MGS rebuild nidtbl. -jay */ } -#endif /* Send target_reg message to MGS */ static int mgc_target_register(struct obd_export *exp, @@ -932,6 +1073,8 @@ static int mgc_target_register(struct obd_export *exp, memcpy(req_mti, mti, sizeof(*req_mti)); ptlrpc_request_set_replen(req); CDEBUG(D_MGC, "register %s\n", mti->mti_svname); + /* Limit how long we will wait for the enqueue to complete */ + req->rq_delay_limit = MGC_TARGET_REG_LIMIT; rc = ptlrpc_queue_wait(req); if (!rc) { @@ -946,44 +1089,28 @@ static int mgc_target_register(struct obd_export *exp, RETURN(rc); } -int mgc_set_info_async(struct obd_export *exp, obd_count keylen, - void *key, obd_count vallen, void *val, - struct ptlrpc_request_set *set) +int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp, + obd_count keylen, void *key, obd_count vallen, + void *val, struct ptlrpc_request_set *set) { - struct obd_import *imp = class_exp2cliimp(exp); int rc = -EINVAL; ENTRY; - /* Try to "recover" the initial connection; i.e. retry */ - if (KEY_IS(KEY_INIT_RECOV)) { - if (vallen != sizeof(int)) - RETURN(-EINVAL); - spin_lock(&imp->imp_lock); - imp->imp_initial_recov = *(int *)val; - spin_unlock(&imp->imp_lock); - CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n", - exp->exp_obd->obd_name, imp->imp_initial_recov); - RETURN(0); - } /* Turn off initial_recov after we try all backup servers once */ if (KEY_IS(KEY_INIT_RECOV_BACKUP)) { + struct obd_import *imp = class_exp2cliimp(exp); int value; if (vallen != sizeof(int)) RETURN(-EINVAL); value = *(int *)val; - spin_lock(&imp->imp_lock); - imp->imp_initial_recov_bk = value > 0; - /* Even after the initial connection, give up all comms if - nobody answers the first time. */ - imp->imp_recon_bk = 1; - spin_unlock(&imp->imp_lock); - CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n", - imp->imp_obd->obd_name, value, imp->imp_initial_recov, + CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n", + imp->imp_obd->obd_name, value, imp->imp_deactive, imp->imp_invalid, imp->imp_replayable, imp->imp_obd->obd_replayable, ptlrpc_import_state_name(imp->imp_state)); /* Resurrect if we previously died */ - if (imp->imp_invalid || value > 1) + if ((imp->imp_state != LUSTRE_IMP_FULL && + imp->imp_state != LUSTRE_IMP_NEW) || value > 1) ptlrpc_reconnect_import(imp); RETURN(0); } @@ -998,27 +1125,21 @@ int mgc_set_info_async(struct obd_export *exp, obd_count keylen, rc = mgc_target_register(exp, mti); RETURN(rc); } - if (KEY_IS(KEY_SET_FS)) { - struct super_block *sb = (struct super_block *)val; - struct lustre_sb_info *lsi; - if (vallen != sizeof(struct super_block)) - RETURN(-EINVAL); - lsi = s2lsi(sb); - rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt); - if (rc) { - CERROR("set_fs got %d\n", rc); - } - RETURN(rc); - } - if (KEY_IS(KEY_CLEAR_FS)) { - if (vallen != 0) - RETURN(-EINVAL); - rc = mgc_fs_cleanup(exp->exp_obd); - if (rc) { - CERROR("clear_fs got %d\n", rc); - } - RETURN(rc); - } + if (KEY_IS(KEY_SET_FS)) { + struct super_block *sb = (struct super_block *)val; + + if (vallen != sizeof(struct super_block)) + RETURN(-EINVAL); + + rc = mgc_fs_setup(exp->exp_obd, sb); + RETURN(rc); + } + if (KEY_IS(KEY_CLEAR_FS)) { + if (vallen != 0) + RETURN(-EINVAL); + rc = mgc_fs_cleanup(exp->exp_obd); + RETURN(rc); + } if (KEY_IS(KEY_SET_INFO)) { struct mgs_send_param *msp; @@ -1073,6 +1194,25 @@ int mgc_set_info_async(struct obd_export *exp, obd_count keylen, RETURN(rc); } +static int mgc_get_info(const struct lu_env *env, struct obd_export *exp, + __u32 keylen, void *key, __u32 *vallen, void *val, + struct lov_stripe_md *unused) +{ + int rc = -EINVAL; + + if (KEY_IS(KEY_CONN_DATA)) { + struct obd_import *imp = class_exp2cliimp(exp); + struct obd_connect_data *data = val; + + if (*vallen == sizeof(*data)) { + *data = imp->imp_connect_data; + rc = 0; + } + } + + return rc; +} + static int mgc_import_event(struct obd_device *obd, struct obd_import *imp, enum obd_import_event event) @@ -1085,6 +1225,8 @@ static int mgc_import_event(struct obd_device *obd, switch (event) { case IMP_EVENT_DISCON: /* MGC imports should not wait for recovery */ + if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV)) + ptlrpc_pinger_ir_down(); break; case IMP_EVENT_INACTIVE: break; @@ -1093,13 +1235,19 @@ static int mgc_import_event(struct obd_device *obd, ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY); break; } - case IMP_EVENT_ACTIVE: - LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name); - /* Clearing obd_no_recov allows us to continue pinging */ - obd->obd_no_recov = 0; - break; + case IMP_EVENT_ACTIVE: + CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name); + /* Clearing obd_no_recov allows us to continue pinging */ + obd->obd_no_recov = 0; + mgc_notify_active(obd); + if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV)) + ptlrpc_pinger_ir_up(); + break; case IMP_EVENT_OCD: break; + case IMP_EVENT_DEACTIVATE: + case IMP_EVENT_ACTIVATE: + break; default: CERROR("Unknown import event %#x\n", event); LBUG(); @@ -1107,195 +1255,520 @@ static int mgc_import_event(struct obd_device *obd, RETURN(rc); } -static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg, - struct obd_device *tgt, int *index) +enum { + CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT), + CONFIG_READ_NRPAGES = 4 +}; + +static int mgc_apply_recover_logs(struct obd_device *mgc, + struct config_llog_data *cld, + __u64 max_version, + void *data, int datalen, bool mne_swab) { - struct llog_ctxt *ctxt; - int rc; + struct config_llog_instance *cfg = &cld->cld_cfg; + struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb); + struct mgs_nidtbl_entry *entry; + struct lustre_cfg *lcfg; + struct lustre_cfg_bufs bufs; + u64 prev_version = 0; + char *inst; + char *buf; + int bufsz; + int pos; + int rc = 0; + int off = 0; ENTRY; - LASSERT(olg == &obd->obd_olg); + LASSERT(cfg->cfg_instance != NULL); + LASSERT(cfg->cfg_sb == cfg->cfg_instance); - rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL, - &llog_lvfs_ops); - if (rc) - RETURN(rc); + OBD_ALLOC(inst, PAGE_CACHE_SIZE); + if (inst == NULL) + RETURN(-ENOMEM); - rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL, - &llog_client_ops); - if (rc == 0) { - ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); - if (!ctxt) { - ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT); - if (ctxt) - llog_cleanup(ctxt); - RETURN(-ENODEV); - } - llog_initiator_connect(ctxt); - llog_ctxt_put(ctxt); + if (!IS_SERVER(lsi)) { + pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance); + if (pos >= PAGE_CACHE_SIZE) { + OBD_FREE(inst, PAGE_CACHE_SIZE); + return -E2BIG; + } } else { - ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT); - if (ctxt) - llog_cleanup(ctxt); + LASSERT(IS_MDT(lsi)); + rc = server_name2svname(lsi->lsi_svname, inst, NULL, + PAGE_CACHE_SIZE); + if (rc) { + OBD_FREE(inst, PAGE_CACHE_SIZE); + RETURN(-EINVAL); + } + pos = strlen(inst); } - RETURN(rc); -} + ++pos; + buf = inst + pos; + bufsz = PAGE_CACHE_SIZE - pos; + + while (datalen > 0) { + int entry_len = sizeof(*entry); + int is_ost; + struct obd_device *obd; + char *obdname; + char *cname; + char *params; + char *uuid; + + rc = -EINVAL; + if (datalen < sizeof(*entry)) + break; -static int mgc_llog_finish(struct obd_device *obd, int count) -{ - struct llog_ctxt *ctxt; - int rc = 0, rc2 = 0; - ENTRY; + entry = (typeof(entry))(data + off); - ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); - if (ctxt) - rc = llog_cleanup(ctxt); + /* sanity check */ + if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */ + break; + if (entry->mne_nid_count == 0) /* at least one nid entry */ + break; + if (entry->mne_nid_size != sizeof(lnet_nid_t)) + break; - ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT); - if (ctxt) - rc2 = llog_cleanup(ctxt); + entry_len += entry->mne_nid_count * entry->mne_nid_size; + if (datalen < entry_len) /* must have entry_len at least */ + break; - if (!rc) - rc = rc2; + /* Keep this swab for normal mixed endian handling. LU-1644 */ + if (mne_swab) + lustre_swab_mgs_nidtbl_entry(entry); + if (entry->mne_length > PAGE_CACHE_SIZE) { + CERROR("MNE too large (%u)\n", entry->mne_length); + break; + } - RETURN(rc); -} + if (entry->mne_length < entry_len) + break; -/* identical to mgs_log_is_empty */ -static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt, - char *name) -{ - struct lvfs_run_ctxt saved; - struct llog_handle *llh; - int rc = 0; + off += entry->mne_length; + datalen -= entry->mne_length; + if (datalen < 0) + break; - push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - rc = llog_create(ctxt, &llh, NULL, name); - if (rc == 0) { - llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL); - rc = llog_get_size(llh); - llog_close(llh); - } - pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - /* header is record 1 */ - return(rc <= 1); -} + if (entry->mne_version > max_version) { + CERROR("entry index(%lld) is over max_index(%lld)\n", + entry->mne_version, max_version); + break; + } -static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec, - void *data) -{ - struct llog_rec_hdr local_rec = *rec; - struct llog_handle *local_llh = (struct llog_handle *)data; - char *cfg_buf = (char*) (rec + 1); - struct lustre_cfg *lcfg; - int rc = 0; - ENTRY; + if (prev_version >= entry->mne_version) { + CERROR("index unsorted, prev %lld, now %lld\n", + prev_version, entry->mne_version); + break; + } + prev_version = entry->mne_version; + + /* + * Write a string with format "nid::instance" to + * lustre//--/import. + */ + + is_ost = entry->mne_type == LDD_F_SV_TYPE_OST; + memset(buf, 0, bufsz); + obdname = buf; + pos = 0; + + /* lustre-OST0001-osc- */ + strcpy(obdname, cld->cld_logname); + cname = strrchr(obdname, '-'); + if (cname == NULL) { + CERROR("mgc %s: invalid logname %s\n", + mgc->obd_name, obdname); + break; + } + + pos = cname - obdname; + obdname[pos] = 0; + pos += sprintf(obdname + pos, "-%s%04x", + is_ost ? "OST" : "MDT", entry->mne_index); + + cname = is_ost ? "osc" : "mdc", + pos += sprintf(obdname + pos, "-%s-%s", cname, inst); + lustre_cfg_bufs_reset(&bufs, obdname); + + /* find the obd by obdname */ + obd = class_name2obd(obdname); + if (obd == NULL) { + CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n", + mgc->obd_name, obdname); + rc = 0; + /* this is a safe race, when the ost is starting up...*/ + continue; + } + + /* osc.import = "connection=::" */ + ++pos; + params = buf + pos; + pos += sprintf(params, "%s.import=%s", cname, "connection="); + uuid = buf + pos; + + down_read(&obd->u.cli.cl_sem); + if (obd->u.cli.cl_import == NULL) { + /* client does not connect to the OST yet */ + up_read(&obd->u.cli.cl_sem); + rc = 0; + continue; + } + + /* TODO: iterate all nids to find one */ + /* find uuid by nid */ + rc = client_import_find_conn(obd->u.cli.cl_import, + entry->u.nids[0], + (struct obd_uuid *)uuid); + up_read(&obd->u.cli.cl_sem); + if (rc < 0) { + CERROR("mgc: cannot find uuid by nid %s\n", + libcfs_nid2str(entry->u.nids[0])); + break; + } + + CDEBUG(D_INFO, "Find uuid %s by nid %s\n", + uuid, libcfs_nid2str(entry->u.nids[0])); + + pos += strlen(uuid); + pos += sprintf(buf + pos, "::%u", entry->mne_instance); + LASSERT(pos < bufsz); - /* Append all records */ - local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail); - rc = llog_write_rec(local_llh, &local_rec, NULL, 0, - (void *)cfg_buf, -1); + lustre_cfg_bufs_set_string(&bufs, 1, params); - lcfg = (struct lustre_cfg *)cfg_buf; - CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n", - rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command, - lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1)); + rc = -ENOMEM; + lcfg = lustre_cfg_new(LCFG_PARAM, &bufs); + if (lcfg == NULL) { + CERROR("mgc: cannot allocate memory\n"); + break; + } + + CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n", + prev_version, max_version, obdname, params); + + rc = class_process_config(lcfg); + lustre_cfg_free(lcfg); + if (rc) + CDEBUG(D_INFO, "process config for %s error %d\n", + obdname, rc); + /* continue, even one with error */ + } + + OBD_FREE(inst, PAGE_CACHE_SIZE); RETURN(rc); } -/* Copy a remote log locally */ -static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt, - struct llog_ctxt *lctxt, char *logname) +/** + * This function is called if this client was notified for target restarting + * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs. + */ +static int mgc_process_recover_log(struct obd_device *obd, + struct config_llog_data *cld) { - struct llog_handle *local_llh, *remote_llh; - struct obd_uuid *uuid; - char *temp_log; - int rc, rc2; + struct ptlrpc_request *req = NULL; + struct config_llog_instance *cfg = &cld->cld_cfg; + struct mgs_config_body *body; + struct mgs_config_res *res; + struct ptlrpc_bulk_desc *desc; + struct page **pages; + int nrpages; + bool eof = true; + bool mne_swab = false; + int i; + int ealen; + int rc; ENTRY; - /* Write new log to a temp name, then vfs_rename over logname - upon successful completion. */ + /* allocate buffer for bulk transfer. + * if this is the first time for this mgs to read logs, + * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs + * once; otherwise, it only reads increment of logs, this should be + * small and CONFIG_READ_NRPAGES will be used. + */ + nrpages = CONFIG_READ_NRPAGES; + if (cfg->cfg_last_idx == 0) /* the first time */ + nrpages = CONFIG_READ_NRPAGES_INIT; + + OBD_ALLOC(pages, sizeof(*pages) * nrpages); + if (pages == NULL) + GOTO(out, rc = -ENOMEM); + + for (i = 0; i < nrpages; i++) { + pages[i] = alloc_page(GFP_IOFS); + if (pages[i] == NULL) + GOTO(out, rc = -ENOMEM); + } - OBD_ALLOC(temp_log, strlen(logname) + 1); - if (!temp_log) - RETURN(-ENOMEM); - sprintf(temp_log, "%sT", logname); +again: + LASSERT(cld_is_recover(cld)); + LASSERT(mutex_is_locked(&cld->cld_lock)); + req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp), + &RQF_MGS_CONFIG_READ); + if (req == NULL) + GOTO(out, rc = -ENOMEM); - /* Make sure there's no old temp log */ - rc = llog_create(lctxt, &local_llh, NULL, temp_log); + rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ); if (rc) GOTO(out, rc); - rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL); - if (rc) - GOTO(out, rc); - rc = llog_destroy(local_llh); - llog_free_handle(local_llh); + + /* pack request */ + body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY); + LASSERT(body != NULL); + LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname)); + if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name)) + >= sizeof(body->mcb_name)) + GOTO(out, rc = -E2BIG); + body->mcb_offset = cfg->cfg_last_idx + 1; + body->mcb_type = cld->cld_type; + body->mcb_bits = PAGE_CACHE_SHIFT; + body->mcb_units = nrpages; + + /* allocate bulk transfer descriptor */ + desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK, + MGS_BULK_PORTAL); + if (desc == NULL) + GOTO(out, rc = -ENOMEM); + + for (i = 0; i < nrpages; i++) + ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE); + + ptlrpc_request_set_replen(req); + rc = ptlrpc_queue_wait(req); if (rc) GOTO(out, rc); - /* open local log */ - rc = llog_create(lctxt, &local_llh, NULL, temp_log); - if (rc) + res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES); + if (res->mcr_size < res->mcr_offset) + GOTO(out, rc = -EINVAL); + + /* always update the index even though it might have errors with + * handling the recover logs */ + cfg->cfg_last_idx = res->mcr_offset; + eof = res->mcr_offset == res->mcr_size; + + CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n", + res->mcr_offset, eof == false); + + ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0); + if (ealen < 0) + GOTO(out, rc = ealen); + + if (ealen > nrpages << PAGE_CACHE_SHIFT) + GOTO(out, rc = -EINVAL); + + if (ealen == 0) { /* no logs transferred */ + if (!eof) + rc = -EINVAL; GOTO(out, rc); + } - /* set the log header uuid for fun */ - OBD_ALLOC_PTR(uuid); - obd_str2uuid(uuid, logname); - rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid); - OBD_FREE_PTR(uuid); - if (rc) - GOTO(out_closel, rc); + mne_swab = !!ptlrpc_rep_need_swab(req); +#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0) + /* This import flag means the server did an extra swab of IR MNE + * records (fixed in LU-1252), reverse it here if needed. LU-1644 */ + if (unlikely(req->rq_import->imp_need_mne_swab)) + mne_swab = !mne_swab; +#else +#warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab" +#endif - /* open remote log */ - rc = llog_create(rctxt, &remote_llh, NULL, logname); - if (rc) - GOTO(out_closel, rc); - rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL); - if (rc) - GOTO(out_closer, rc); - - /* Copy remote log */ - rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL); - -out_closer: - rc2 = llog_close(remote_llh); - if (!rc) - rc = rc2; -out_closel: - rc2 = llog_close(local_llh); - if (!rc) - rc = rc2; - - /* We've copied the remote log to the local temp log, now - replace the old local log with the temp log. */ - if (!rc) { - struct client_obd *cli = &obd->u.cli; - LASSERT(cli); - LASSERT(cli->cl_mgc_configs_dir); - rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt, - temp_log, logname); + for (i = 0; i < nrpages && ealen > 0; i++) { + int rc2; + void *ptr; + + ptr = kmap(pages[i]); + rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr, + min_t(int, ealen, PAGE_CACHE_SIZE), + mne_swab); + kunmap(pages[i]); + if (rc2 < 0) { + CWARN("Process recover log %s error %d\n", + cld->cld_logname, rc2); + break; + } + + ealen -= PAGE_CACHE_SIZE; } - CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc); + out: - if (rc) - CERROR("Failed to copy remote log %s (%d)\n", logname, rc); - OBD_FREE(temp_log, strlen(logname) + 1); - RETURN(rc); + if (req) + ptlrpc_req_finished(req); + + if (rc == 0 && !eof) + goto again; + + if (pages) { + for (i = 0; i < nrpages; i++) { + if (pages[i] == NULL) + break; + __free_page(pages[i]); + } + OBD_FREE(pages, sizeof(*pages) * nrpages); + } + return rc; } -/* Get a config log from the MGS and process it. - This func is called for both clients and servers. */ -int mgc_process_log(struct obd_device *mgc, - struct config_llog_data *cld) +/* Copy a remote log locally */ +static int mgc_llog_local_copy(const struct lu_env *env, + struct obd_device *obd, + struct llog_ctxt *rctxt, + struct llog_ctxt *lctxt, char *logname) { - struct llog_ctxt *ctxt, *lctxt; - struct lustre_handle lockh; - struct client_obd *cli = &mgc->u.cli; - struct lvfs_run_ctxt saved; - struct lustre_sb_info *lsi = NULL; - int rc = 0, rcl, flags = 0, must_pop = 0; + char *temp_log; + int rc; + + ENTRY; + + /* + * - copy it to backup using llog_backup() + * - copy remote llog to logname using llog_backup() + * - if failed then move bakup to logname again + */ + + OBD_ALLOC(temp_log, strlen(logname) + 1); + if (!temp_log) + RETURN(-ENOMEM); + sprintf(temp_log, "%sT", logname); + + /* make a copy of local llog at first */ + rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log); + if (rc < 0 && rc != -ENOENT) + GOTO(out, rc); + /* copy remote llog to the local copy */ + rc = llog_backup(env, obd, rctxt, lctxt, logname, logname); + if (rc == -ENOENT) { + /* no remote llog, delete local one too */ + llog_erase(env, lctxt, NULL, logname); + } else if (rc < 0) { + /* error during backup, get local one back from the copy */ + llog_backup(env, obd, lctxt, lctxt, temp_log, logname); +out: + CERROR("%s: failed to copy remote log %s: rc = %d\n", + obd->obd_name, logname, rc); + } + llog_erase(env, lctxt, NULL, temp_log); + OBD_FREE(temp_log, strlen(logname) + 1); + return rc; +} + +/* local_only means it cannot get remote llogs */ +static int mgc_process_cfg_log(struct obd_device *mgc, + struct config_llog_data *cld, int local_only) +{ + struct llog_ctxt *ctxt, *lctxt = NULL; + struct client_obd *cli = &mgc->u.cli; + struct lustre_sb_info *lsi = NULL; + int rc = 0; + bool sptlrpc_started = false; + struct lu_env *env; + + ENTRY; + + LASSERT(cld); + LASSERT(mutex_is_locked(&cld->cld_lock)); + + /* + * local copy of sptlrpc log is controlled elsewhere, don't try to + * read it up here. + */ + if (cld_is_sptlrpc(cld) && local_only) + RETURN(0); + + if (cld->cld_cfg.cfg_sb) + lsi = s2lsi(cld->cld_cfg.cfg_sb); + + OBD_ALLOC_PTR(env); + if (env == NULL) + RETURN(-ENOMEM); + + rc = lu_env_init(env, LCT_MG_THREAD); + if (rc) + GOTO(out_free, rc); + + ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT); + LASSERT(ctxt); + + lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT); + + /* Copy the setup log locally if we can. Don't mess around if we're + * running an MGS though (logs are already local). */ + if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) && + cli->cl_mgc_configs_dir != NULL && + lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) == + lsi->lsi_dt_dev) { + if (!local_only) + /* Only try to copy log if we have the lock. */ + rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt, + cld->cld_logname); + if (local_only || rc) { + if (llog_is_empty(env, lctxt, cld->cld_logname)) { + LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS " + "log %s and no local copy." + "\n", cld->cld_logname); + GOTO(out_pop, rc = -ENOTCONN); + } + CDEBUG(D_MGC, "Failed to get MGS log %s, using local " + "copy for now, will try to update later.\n", + cld->cld_logname); + } + /* Now, whether we copied or not, start using the local llog. + * If we failed to copy, we'll start using whatever the old + * log has. */ + llog_ctxt_put(ctxt); + ctxt = lctxt; + lctxt = NULL; + } else { + if (local_only) /* no local log at client side */ + GOTO(out_pop, rc = -EIO); + } + + if (cld_is_sptlrpc(cld)) { + sptlrpc_conf_log_update_begin(cld->cld_logname); + sptlrpc_started = true; + } + + /* logname and instance info should be the same, so use our + * copy of the instance for the update. The cfg_last_idx will + * be updated here. */ + rc = class_config_parse_llog(env, ctxt, cld->cld_logname, + &cld->cld_cfg); + EXIT; + +out_pop: + __llog_ctxt_put(env, ctxt); + if (lctxt) + __llog_ctxt_put(env, lctxt); + + /* + * update settings on existing OBDs. doing it inside + * of llog_process_lock so no device is attaching/detaching + * in parallel. + * the logname must be -sptlrpc + */ + if (sptlrpc_started) { + LASSERT(cld_is_sptlrpc(cld)); + sptlrpc_conf_log_update_end(cld->cld_logname); + class_notify_sptlrpc_conf(cld->cld_logname, + strlen(cld->cld_logname) - + strlen("-sptlrpc")); + } + + lu_env_fini(env); +out_free: + OBD_FREE_PTR(env); + return rc; +} + +/** Get a config log from the MGS and process it. + * This func is called for both clients and servers. + * Copy the log locally before parsing it if appropriate (non-MGS server) + */ +int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld) +{ + struct lustre_handle lockh = { 0 }; + __u64 flags = LDLM_FL_NO_LRU; + int rc = 0, rcl; ENTRY; LASSERT(cld); @@ -1304,98 +1777,49 @@ int mgc_process_log(struct obd_device *mgc, sounds like badness. It actually might be fine, as long as we're not trying to update from the same log simultaneously (in which case we should use a per-log sem.) */ - down(&llog_process_lock); - - if (cld->cld_stopping) { - up(&llog_process_lock); + mutex_lock(&cld->cld_lock); + if (cld->cld_stopping) { + mutex_unlock(&cld->cld_lock); RETURN(0); } OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20); - if (cld->cld_cfg.cfg_sb) - lsi = s2lsi(cld->cld_cfg.cfg_sb); - - CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname, + CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname, cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1); - ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT); - if (!ctxt) { - CERROR("missing llog context\n"); - up(&llog_process_lock); - RETURN(-EINVAL); - } - /* Get the cfg lock on the llog */ rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL, LCK_CR, &flags, NULL, NULL, NULL, cld, 0, NULL, &lockh); - if (rcl) + if (rcl == 0) { + /* Get the cld, it will be released in mgc_blocking_ast. */ + config_log_get(cld); + rc = ldlm_lock_set_data(&lockh, (void *)cld); + LASSERT(rc == 0); + } else { CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl); - lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT); + /* mark cld_lostlock so that it will requeue + * after MGC becomes available. */ + cld->cld_lostlock = 1; + /* Get extra reference, it will be put in requeue thread */ + config_log_get(cld); + } - /* - * local copy of sptlrpc log is controlled elsewhere, don't try to - * read it up here. - */ - if (rcl && cld->cld_is_sptlrpc) - goto out_pop; - - /* Copy the setup log locally if we can. Don't mess around if we're - running an MGS though (logs are already local). */ - if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) && - (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) && - !IS_MGS(lsi->lsi_ldd)) { - push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL); - must_pop++; + + if (cld_is_recover(cld)) { + rc = 0; /* this is not a fatal error for recover log */ if (rcl == 0) - /* Only try to copy log if we have the lock. */ - rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname); - if (rcl || rc) { - if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) { - LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS " - "log %s and no local copy." - "\n", cld->cld_logname); - GOTO(out_pop, rc = -ENOTCONN); - } - CDEBUG(D_MGC, "Failed to get MGS log %s, using local " - "copy for now, will try to update later.\n", - cld->cld_logname); - } - /* Now, whether we copied or not, start using the local llog. - If we failed to copy, we'll start using whatever the old - log has. */ - llog_ctxt_put(ctxt); - ctxt = lctxt; + rc = mgc_process_recover_log(mgc, cld); + } else { + rc = mgc_process_cfg_log(mgc, cld, rcl != 0); } - if (cld->cld_is_sptlrpc) - sptlrpc_conf_log_update_begin(cld->cld_logname); - - /* logname and instance info should be the same, so use our - copy of the instance for the update. The cfg_last_idx will - be updated here. */ - rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg); -out_pop: - llog_ctxt_put(ctxt); - if (ctxt != lctxt) - llog_ctxt_put(lctxt); - if (must_pop) - pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL); + CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n", + mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc); - /* - * update settings on existing OBDs. doing it inside - * of llog_process_lock so no device is attaching/detaching - * in parallel. - * the logname must be -sptlrpc - */ - if (cld->cld_is_sptlrpc && rcl == 0) { - sptlrpc_conf_log_update_end(cld->cld_logname); - class_notify_sptlrpc_conf(cld->cld_logname, - strlen(cld->cld_logname) - - strlen("-sptlrpc")); - } + mutex_unlock(&cld->cld_lock); /* Now drop the lock so MGS can revoke it */ if (!rcl) { @@ -1405,24 +1829,25 @@ out_pop: CERROR("Can't drop cfg lock: %d\n", rcl); } - CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n", - mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc); - - up(&llog_process_lock); - RETURN(rc); } + +/** Called from lustre_process_log. + * LCFG_LOG_START gets the config log from the MGS, processes it to start + * any services, and adds it to the list logs to watch (follow). + */ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf) { struct lustre_cfg *lcfg = buf; - int cmd; + struct config_llog_instance *cfg = NULL; + char *logname; int rc = 0; ENTRY; - switch(cmd = lcfg->lcfg_command) { + switch(lcfg->lcfg_command) { case LCFG_LOV_ADD_OBD: { - /* Add any new target, not just osts */ + /* Overloading this cfg command: register a new target */ struct mgs_target_info *mti; if (LUSTRE_CFG_BUFLEN(lcfg, 1) != @@ -1436,8 +1861,7 @@ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf) break; } case LCFG_LOV_DEL_OBD: - /* Remove target from the fs? */ - /* FIXME */ + /* Unregister has no meaning at the moment. */ CERROR("lov_del_obd unimplemented\n"); rc = -ENOSYS; break; @@ -1447,9 +1871,9 @@ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf) } case LCFG_LOG_START: { struct config_llog_data *cld; - struct config_llog_instance *cfg; struct super_block *sb; - char *logname = lustre_cfg_string(lcfg, 1); + + logname = lustre_cfg_string(lcfg, 1); cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2); sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3); @@ -1461,8 +1885,8 @@ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf) if (rc) break; cld = config_log_find(logname, cfg); - if (IS_ERR(cld)) { - rc = PTR_ERR(cld); + if (cld == NULL) { + rc = -ENOENT; break; } @@ -1472,13 +1896,25 @@ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf) cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146; rc = mgc_process_log(obd, cld); + if (rc == 0 && cld->cld_recover != NULL) { + if (OCD_HAS_FLAG(&obd->u.cli.cl_import-> + imp_connect_data, IMP_RECOV)) { + rc = mgc_process_log(obd, cld->cld_recover); + } else { + struct config_llog_data *cir = cld->cld_recover; + cld->cld_recover = NULL; + config_log_put(cir); + } + if (rc) + CERROR("Cannot process recover llog %d\n", rc); + } config_log_put(cld); break; } case LCFG_LOG_END: { - struct config_llog_instance *cfg = NULL; - char *logname = lustre_cfg_string(lcfg, 1); + logname = lustre_cfg_string(lcfg, 1); + if (lcfg->lcfg_bufcount >= 2) cfg = (struct config_llog_instance *)lustre_cfg_buf( lcfg, 2); @@ -1508,9 +1944,8 @@ struct obd_ops mgc_obd_ops = { .o_cancel = mgc_cancel, //.o_iocontrol = mgc_iocontrol, .o_set_info_async = mgc_set_info_async, + .o_get_info = mgc_get_info, .o_import_event = mgc_import_event, - .o_llog_init = mgc_llog_init, - .o_llog_finish = mgc_llog_finish, .o_process_config = mgc_process_config, };