X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fobdfilter%2Ffilter.c;h=422db94f179872e7d70fc717c5ad35794f839d4a;hb=11c2c0ec77125041e9c8143a80e7e51aede653ea;hp=9dc6584449488afcb3676f812b2c3e9acd92e0c1;hpb=774512ea1e17efb99713e29f4dff03c71843b211;p=fs%2Flustre-release.git diff --git a/lustre/obdfilter/filter.c b/lustre/obdfilter/filter.c index 9dc6584..422db94 100644 --- a/lustre/obdfilter/filter.c +++ b/lustre/obdfilter/filter.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. @@ -53,9 +51,6 @@ #define DEBUG_SUBSYSTEM S_FILTER -#ifndef AUTOCONF_INCLUDED -#include -#endif #include #include #include @@ -1850,15 +1845,16 @@ static int filter_intent_policy(struct ldlm_namespace *ns, * at the OST layer there are only (potentially) multiple obd_device of type * unknown at the time of OST thread creation. * - * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool - * field). This array has size OST_MAX_THREADS, so that each OST thread uses - * it's very own iobuf. + * We create a cfs_hash for struct filter_obd (->fo_iobuf_hash field) on + * initializing, each OST thread will create it's own iobuf on the first + * access and insert it into ->fo_iobuf_hash with thread ID as key, + * so the iobuf can be found again by thread ID. * * Functions below * - * filter_kiobuf_pool_init() + * filter_iobuf_pool_init() * - * filter_kiobuf_pool_done() + * filter_iobuf_pool_done() * * filter_iobuf_get() * @@ -1871,21 +1867,13 @@ static int filter_intent_policy(struct ldlm_namespace *ns, */ static void filter_iobuf_pool_done(struct filter_obd *filter) { - struct filter_iobuf **pool; - int i; + ENTRY; - ENTRY; - - pool = filter->fo_iobuf_pool; - if (pool != NULL) { - for (i = 0; i < filter->fo_iobuf_count; ++ i) { - if (pool[i] != NULL) - filter_free_iobuf(pool[i]); - } - OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]); - filter->fo_iobuf_pool = NULL; - } - EXIT; + if (filter->fo_iobuf_hash != NULL) { + cfs_hash_putref(filter->fo_iobuf_hash); + filter->fo_iobuf_hash = NULL; + } + EXIT; } static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial) @@ -1912,50 +1900,126 @@ static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial) return 0; } -/* - * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write(). - */ -static int filter_iobuf_pool_init(struct filter_obd *filter) +static unsigned +filter_iobuf_hop_hash(cfs_hash_t *hs, const void *key, unsigned mask) { - void **pool; + __u64 val = *((__u64 *)key); - ENTRY; + return cfs_hash_long(val, hs->hs_cur_bits); +} +static void * +filter_iobuf_hop_key(cfs_hlist_node_t *hnode) +{ + struct filter_iobuf *pool; - OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool), - GFP_KERNEL); - if (filter->fo_iobuf_pool == NULL) - RETURN(-ENOMEM); + pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist); + return &pool->dr_hkey; +} - filter->fo_iobuf_count = OSS_THREADS_MAX; +static int +filter_iobuf_hop_keycmp(const void *key, cfs_hlist_node_t *hnode) +{ + struct filter_iobuf *pool; - RETURN(0); + pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist); + return pool->dr_hkey == *((__u64 *)key); } -/* Return iobuf allocated for @thread_id. We don't know in advance how - * many threads there will be so we allocate a large empty array and only - * fill in those slots that are actually in use. - * If we haven't allocated a pool entry for this thread before, do so now. */ -void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti) +static void * +filter_iobuf_hop_object(cfs_hlist_node_t *hnode) { - int thread_id = (oti && oti->oti_thread) ? - oti->oti_thread->t_id : -1; - struct filter_iobuf *pool = NULL; - struct filter_iobuf **pool_place = NULL; + return cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist); +} - if (thread_id >= 0) { - LASSERT(thread_id < filter->fo_iobuf_count); - pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]); - } +static void +filter_iobuf_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode) +{ + /* dummy, required by cfs_hash */ +} - if (unlikely(pool == NULL)) { - pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE, - PTLRPC_MAX_BRW_PAGES); - if (pool_place != NULL) - *pool_place = pool; - } +static void +filter_iobuf_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode) +{ + /* dummy, required by cfs_hash */ +} + +static void +filter_iobuf_hop_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode) +{ + struct filter_iobuf *pool; + + pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist); + filter_free_iobuf(pool); +} + +static struct cfs_hash_ops filter_iobuf_hops = { + .hs_hash = filter_iobuf_hop_hash, + .hs_key = filter_iobuf_hop_key, + .hs_keycmp = filter_iobuf_hop_keycmp, + .hs_object = filter_iobuf_hop_object, + .hs_get = filter_iobuf_hop_get, + .hs_put_locked = filter_iobuf_hop_put_locked, + .hs_exit = filter_iobuf_hop_exit +}; + +#define FILTER_IOBUF_HASH_BITS 9 +#define FILTER_IOBUF_HBKT_BITS 4 + +/* + * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write(). + */ +static int filter_iobuf_pool_init(struct filter_obd *filter) +{ + filter->fo_iobuf_hash = cfs_hash_create("filter_iobuf", + FILTER_IOBUF_HASH_BITS, + FILTER_IOBUF_HASH_BITS, + FILTER_IOBUF_HBKT_BITS, 0, + CFS_HASH_MIN_THETA, + CFS_HASH_MAX_THETA, + &filter_iobuf_hops, + CFS_HASH_RW_BKTLOCK | + CFS_HASH_NO_ITEMREF); + + return filter->fo_iobuf_hash != NULL ? 0 : -ENOMEM; +} - return pool; +/* Return iobuf allocated for @thread_id. + * If we haven't allocated a pool entry for this thread before, do so now and + * insert it into fo_iobuf_hash, otherwise we can find it from fo_iobuf_hash */ +void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti) +{ + struct filter_iobuf *pool = NULL; + __u64 key = 0; + int thread_id; + int rc; + + thread_id = (oti && oti->oti_thread) ? oti->oti_thread->t_id : -1; + if (thread_id >= 0) { + struct ptlrpc_service_part *svcpt; + + svcpt = oti->oti_thread->t_svcpt; + LASSERT(svcpt != NULL); + + key = (__u64)(svcpt->scp_cpt) << 32 | thread_id; + pool = cfs_hash_lookup(filter->fo_iobuf_hash, &key); + if (pool != NULL) + return pool; + } + + pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE, PTLRPC_MAX_BRW_PAGES); + if (pool == NULL) + return NULL; + + if (thread_id >= 0) { + pool->dr_hkey = key; + rc = cfs_hash_add_unique(filter->fo_iobuf_hash, + &key, &pool->dr_hlist); + /* ptlrpc service thould guarantee thread ID is unique */ + LASSERT(rc != -EALREADY); + } + + return pool; } /* mount the file system (secretly). lustre_cfg parameters are: @@ -2226,10 +2290,15 @@ static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg) GOTO(free_obd_stats, rc); } + rc = lprocfs_job_stats_init(obd, LPROC_FILTER_STATS_LAST, + filter_stats_counter_init); + if (rc) + GOTO(remove_entry_clear, rc); + /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */ OBD_PAGE_ALLOC(page, CFS_ALLOC_STD); if (!page) - GOTO(remove_entry_clear, rc = -ENOMEM); + GOTO(job_stats_fini, rc = -ENOMEM); addr = (unsigned long)cfs_page_address(page); clear_page((void *)addr); memcpy((void *)addr, lustre_cfg_buf(lcfg, 4), @@ -2239,11 +2308,13 @@ static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg) if (rc) { CERROR("%s: filter_common_setup failed: %d.\n", obd->obd_name, rc); - GOTO(remove_entry_clear, rc); + GOTO(job_stats_fini, rc); } RETURN(0); +job_stats_fini: + lprocfs_job_stats_fini(obd); remove_entry_clear: lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry); free_obd_stats: @@ -2403,7 +2474,7 @@ static int filter_llog_finish(struct obd_device *obd, int count) * We actually do sync in disconnect time, but disconnect * may not come being marked rq_no_resend = 1. */ - llog_sync(ctxt, NULL); + llog_sync(ctxt, NULL, OBD_LLOG_FL_EXIT); /* * Balance class_import_get() in llog_receptor_accept(). @@ -2415,16 +2486,16 @@ static int filter_llog_finish(struct obd_device *obd, int count) class_import_put(ctxt->loc_imp); ctxt->loc_imp = NULL; } + + if (filter->fo_lcm) { + llog_recov_thread_fini(filter->fo_lcm, obd->obd_force); + filter->fo_lcm = NULL; + } + cfs_mutex_unlock(&ctxt->loc_mutex); llog_ctxt_put(ctxt); } - if (filter->fo_lcm) { - cfs_mutex_lock(&ctxt->loc_mutex); - llog_recov_thread_fini(filter->fo_lcm, obd->obd_force); - filter->fo_lcm = NULL; - cfs_mutex_unlock(&ctxt->loc_mutex); - } RETURN(filter_olg_fini(&obd->obd_olg)); } @@ -2550,9 +2621,9 @@ static int filter_llog_connect(struct obd_export *exp, LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n", body->lgdc_ctxt_idx); - CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n", - obd->obd_name, body->lgdc_logid.lgl_oid, - body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen); + CDEBUG(D_HA, "%s: Recovery from log "LPX64"/"LPX64":%x\n", + obd->obd_name, body->lgdc_logid.lgl_oid, + body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen); cfs_spin_lock(&obd->u.filter.fo_flags_lock); obd->u.filter.fo_mds_ost_sync = 1; @@ -2620,6 +2691,7 @@ static int filter_precleanup(struct obd_device *obd, obd_zombie_barrier(); rc = filter_llog_preclean(obd); + lprocfs_job_stats_fini(obd); lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry); lprocfs_free_per_client_stats(obd); lprocfs_obd_cleanup(obd); @@ -2695,8 +2767,8 @@ static int filter_connect_internal(struct obd_export *exp, /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */ if (data->ocd_connect_flags & OBD_CONNECT_MDS) - CWARN("%s: Received MDS connection for group %u\n", - exp->exp_obd->obd_name, data->ocd_group); + CDEBUG(D_HA, "%s: Received MDS connection for group %u\n", + exp->exp_obd->obd_name, data->ocd_group); else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) RETURN(-EPROTO); @@ -2771,11 +2843,9 @@ static int filter_connect_internal(struct obd_export *exp, /* The client set in ocd_cksum_types the checksum types it * supports. We have to mask off the algorithms that we don't * support */ - data->ocd_cksum_types &= cksum_types_supported(); + data->ocd_cksum_types &= cksum_types_supported_server(); - /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */ - if (unlikely(data->ocd_cksum_types == 0)) - data->ocd_cksum_types = OBD_CKSUM_CRC32; + /* 1.6.4 clients are not supported any more */ CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return " "%x\n", exp->exp_obd->obd_name, @@ -3060,7 +3130,7 @@ static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp) ctxt = llog_group_get_ctxt(olg_min, LLOG_MDS_OST_REPL_CTXT); if (ctxt) { - err = llog_sync(ctxt, olg_min->olg_exp); + err = llog_sync(ctxt, olg_min->olg_exp, 0); llog_ctxt_put(ctxt); if (err) { CERROR("error flushing logs to MDS: " @@ -3116,7 +3186,7 @@ static void filter_revimp_update(struct obd_export *exp) EXIT; } -static int filter_ping(struct obd_export *exp) +static int filter_ping(const struct lu_env *env, struct obd_export *exp) { filter_fmd_expire(exp); return 0; @@ -3162,7 +3232,8 @@ struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid, return dchild; } -static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo) +static int filter_getattr(const struct lu_env *env, struct obd_export *exp, + struct obd_info *oinfo) { struct dentry *dentry = NULL; struct obd_device *obd; @@ -3419,8 +3490,8 @@ out_unlock: } /* this is called from filter_truncate() until we have filter_punch() */ -int filter_setattr(struct obd_export *exp, struct obd_info *oinfo, - struct obd_trans_info *oti) +int filter_setattr(const struct lu_env *env, struct obd_export *exp, + struct obd_info *oinfo, struct obd_trans_info *oti) { struct obdo *oa = oinfo->oi_oa; struct lustre_capa *capa = oinfo_capa(oinfo); @@ -3514,6 +3585,8 @@ int filter_setattr(struct obd_export *exp, struct obd_info *oinfo, obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID); + filter_counter_incr(exp, LPROC_FILTER_STATS_SETATTR, + oti ? oti->oti_jobid : NULL, 1); EXIT; out_unlock: f_dput(dentry); @@ -3611,7 +3684,7 @@ static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa, for (id = last; id > oa->o_id; id--) { doa.o_id = id; - rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL); + rc = filter_destroy(NULL, exp, &doa, NULL, NULL, NULL, NULL); if (rc && rc != -ENOENT) /* this is pretty fatal... */ CEMERG("error destroying precreate objid "LPU64": %d\n", id, rc); @@ -3680,7 +3753,7 @@ static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa, CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n", last, diff); - if (-diff > OST_MAX_PRECREATE) { + if (-diff > (OST_MAX_PRECREATE * 3) / 2) { CERROR("%s: ignoring bogus orphan destroy request: " "obdid "LPU64" last_id "LPU64"\n", obd->obd_name, oa->o_id, last); @@ -3731,9 +3804,10 @@ out: return rc; } -static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs, - __u64 max_age, __u32 flags) +static int filter_statfs(const struct lu_env *env, struct obd_export *exp, + struct obd_statfs *osfs, __u64 max_age, __u32 flags) { + struct obd_device *obd = class_exp2obd(exp); struct filter_obd *filter = &obd->u.filter; int blockbits = obd->u.obt.obt_sb->s_blocksize_bits; struct lr_server_data *lsd = class_server_data(obd); @@ -3857,7 +3931,7 @@ static int filter_precreate(struct obd_device *obd, struct obdo *oa, OBD_ALLOC(osfs, sizeof(*osfs)); if (osfs == NULL) RETURN(-ENOMEM); - rc = filter_statfs(obd, osfs, + rc = filter_statfs(NULL, obd->obd_self_export, osfs, cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS), 0); if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) { @@ -4059,8 +4133,9 @@ set_last_id: RETURN(rc); } -int filter_create(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md **ea, struct obd_trans_info *oti) +int filter_create(const struct lu_env *env, struct obd_export *exp, + struct obdo *oa, struct lov_stripe_md **ea, + struct obd_trans_info *oti) { struct obd_device *obd = exp->exp_obd; struct filter_export_data *fed; @@ -4139,9 +4214,10 @@ int filter_create(struct obd_export *exp, struct obdo *oa, RETURN(rc); } -int filter_destroy(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *md, struct obd_trans_info *oti, - struct obd_export *md_exp, void *capa) +int filter_destroy(const struct lu_env *env, struct obd_export *exp, + struct obdo *oa, struct lov_stripe_md *md, + struct obd_trans_info *oti, struct obd_export *md_exp, + void *capa) { unsigned int qcids[MAXQUOTAS] = {0, 0}; struct obd_device *obd; @@ -4335,8 +4411,8 @@ cleanup: } /* NB start and end are used for punch, but not truncate */ -static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo, - struct obd_trans_info *oti, +static int filter_truncate(const struct lu_env *env, struct obd_export *exp, + struct obd_info *oinfo, struct obd_trans_info *oti, struct ptlrpc_request_set *rqset) { int rc; @@ -4353,12 +4429,12 @@ static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo, oinfo->oi_policy.l_extent.start); oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start; - rc = filter_setattr(exp, oinfo, oti); + rc = filter_setattr(env, exp, oinfo, oti); RETURN(rc); } -static int filter_sync(struct obd_export *exp, struct obd_info *oinfo, - obd_off start, obd_off end, +static int filter_sync(const struct lu_env *env, struct obd_export *exp, + struct obd_info *oinfo, obd_off start, obd_off end, struct ptlrpc_request_set *set) { struct lvfs_run_ctxt saved; @@ -4410,12 +4486,13 @@ static int filter_sync(struct obd_export *exp, struct obd_info *oinfo, pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL); + filter_counter_incr(exp, LPROC_FILTER_STATS_SYNC, oinfo->oi_jobid, 1); f_dput(dentry); RETURN(rc); } -static int filter_get_info(struct obd_export *exp, __u32 keylen, - void *key, __u32 *vallen, void *val, +static int filter_get_info(const struct lu_env *env, struct obd_export *exp, + __u32 keylen, void *key, __u32 *vallen, void *val, struct lov_stripe_md *lsm) { struct obd_device *obd; @@ -4571,7 +4648,8 @@ out: RETURN(rc); } -static int filter_set_info_async(struct obd_export *exp, __u32 keylen, +static int filter_set_info_async(const struct lu_env *env, + struct obd_export *exp, __u32 keylen, void *key, __u32 vallen, void *val, struct ptlrpc_request_set *set) { @@ -4680,7 +4758,7 @@ int filter_iocontrol(unsigned int cmd, struct obd_export *exp, RETURN(0); } -static int filter_health_check(struct obd_device *obd) +static int filter_health_check(const struct lu_env *env, struct obd_device *obd) { #ifdef USE_HEALTH_CHECK_WRITE struct filter_obd *filter = &obd->u.filter;