From d234e2cf5f557eadd4b65758ca76f783fb6ed9a1 Mon Sep 17 00:00:00 2001 From: NeilBrown Date: Tue, 17 Sep 2019 15:33:09 -0400 Subject: [PATCH] LU-8130 ldlm: separate buckets from ldlm hash table ldlm maintains a per-namespace hashtable of resources. With these hash tables it stores per-bucket 'struct adaptive_timeout' structures. Presumably having a single struct for the whole table results in too much contention while having one per resource results in very little adaption. A future patch will change ldlm to use rhashtable which does not support per-bucket data, so we need to manage the data separately. There is no need for the multiple adaptive_timeout to align with the hash chains, and trying to do this has resulted in a rather complex hash function. The purpose of ldlm_res_hop_fid_hash() appears to be to keep resources with the same fid in the same hash bucket, so they use the same adaptive timeout. However it fails at doing this because it puts the fid-specific bits in the wrong part of the hash. If that is not the purpose, then I can see no point to the complexitiy. This patch creates a completely separate array of adaptive timeouts (and other less interesting data) and uses a hash of the fid to index that, meaning that a simple hash can be used for the hash table. In the previous code, two namespace uses the same value for nsd_all_bits and nsd_bkt_bits. This results in zero bits being used to choose a bucket - so there is only one bucket. This looks odd and would confuse hash_32(), so I've adjusted the numbers so there is always at least 1 bit (2 buckets). Change-Id: Ifab1b48b35b4a9a56610340556875901ad3804b2 Signed-off-by: NeilBrown Reviewed-on: https://review.whamcloud.com/36218 Reviewed-by: Neil Brown Tested-by: jenkins Tested-by: Maloo Reviewed-by: Shaun Tancheff Reviewed-by: Oleg Drokin --- lustre/include/lustre_dlm.h | 2 ++ lustre/ldlm/ldlm_resource.c | 56 +++++++++++++++++++-------------------------- 2 files changed, 26 insertions(+), 32 deletions(-) diff --git a/lustre/include/lustre_dlm.h b/lustre/include/lustre_dlm.h index 181518d..7534486 100644 --- a/lustre/include/lustre_dlm.h +++ b/lustre/include/lustre_dlm.h @@ -384,6 +384,8 @@ struct ldlm_namespace { /** Resource hash table for namespace. */ struct cfs_hash *ns_rs_hash; + struct ldlm_ns_bucket *ns_rs_buckets; + unsigned int ns_bucket_bits; /** serialize */ spinlock_t ns_lock; diff --git a/lustre/ldlm/ldlm_resource.c b/lustre/ldlm/ldlm_resource.c index 24c9e56..9275227 100644 --- a/lustre/ldlm/ldlm_resource.c +++ b/lustre/ldlm/ldlm_resource.c @@ -768,10 +768,8 @@ static unsigned ldlm_res_hop_hash(struct cfs_hash *hs, return val & mask; } -static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs, - const void *key, unsigned mask) +static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits) { - const struct ldlm_res_id *id = key; struct lu_fid fid; __u32 hash; __u32 val; @@ -784,18 +782,11 @@ static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs, hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */ if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) { val = id->name[LUSTRE_RES_ID_HSH_OFF]; - hash += (val >> 5) + (val << 11); } else { val = fid_oid(&fid); } - hash = hash_long(hash, hs->hs_bkt_bits); - /* give me another random factor */ - hash -= hash_long((unsigned long)hs, val % 11 + 3); - - hash <<= hs->hs_cur_bits - hs->hs_bkt_bits; - hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1); - - return hash & mask; + hash += (val >> 5) + (val << 11); + return cfs_hash_32(hash, bits); } static void *ldlm_res_hop_key(struct hlist_node *hnode) @@ -847,16 +838,6 @@ static struct cfs_hash_ops ldlm_ns_hash_ops = { .hs_put = ldlm_res_hop_put }; -static struct cfs_hash_ops ldlm_ns_fid_hash_ops = { - .hs_hash = ldlm_res_hop_fid_hash, - .hs_key = ldlm_res_hop_key, - .hs_keycmp = ldlm_res_hop_keycmp, - .hs_keycpy = NULL, - .hs_object = ldlm_res_hop_object, - .hs_get = ldlm_res_hop_get_locked, - .hs_put = ldlm_res_hop_put -}; - typedef struct ldlm_ns_hash_def { enum ldlm_ns_type nsd_type; /** hash bucket bits */ @@ -873,13 +854,13 @@ static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] = .nsd_type = LDLM_NS_TYPE_MDC, .nsd_bkt_bits = 11, .nsd_all_bits = 16, - .nsd_hops = &ldlm_ns_fid_hash_ops, + .nsd_hops = &ldlm_ns_hash_ops, }, { .nsd_type = LDLM_NS_TYPE_MDT, .nsd_bkt_bits = 14, .nsd_all_bits = 21, - .nsd_hops = &ldlm_ns_fid_hash_ops, + .nsd_hops = &ldlm_ns_hash_ops, }, { .nsd_type = LDLM_NS_TYPE_OSC, @@ -895,13 +876,13 @@ static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] = }, { .nsd_type = LDLM_NS_TYPE_MGC, - .nsd_bkt_bits = 4, + .nsd_bkt_bits = 3, .nsd_all_bits = 4, .nsd_hops = &ldlm_ns_hash_ops, }, { .nsd_type = LDLM_NS_TYPE_MGT, - .nsd_bkt_bits = 4, + .nsd_bkt_bits = 3, .nsd_all_bits = 4, .nsd_hops = &ldlm_ns_hash_ops, }, @@ -919,9 +900,7 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, enum ldlm_ns_type ns_type) { struct ldlm_namespace *ns = NULL; - struct ldlm_ns_bucket *nsb; struct ldlm_ns_hash_def *nsd; - struct cfs_hash_bd bd; int idx; int rc; @@ -951,7 +930,7 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, ns->ns_rs_hash = cfs_hash_create(name, nsd->nsd_all_bits, nsd->nsd_all_bits, - nsd->nsd_bkt_bits, sizeof(*nsb), + nsd->nsd_bkt_bits, 0, CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA, nsd->nsd_hops, @@ -962,8 +941,15 @@ struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, if (ns->ns_rs_hash == NULL) GOTO(out_ns, NULL); - cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) { - nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); + ns->ns_bucket_bits = nsd->nsd_all_bits - nsd->nsd_bkt_bits; + OBD_ALLOC_LARGE(ns->ns_rs_buckets, + BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0])); + if (!ns->ns_rs_buckets) + goto out_hash; + + for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) { + struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx]; + at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0); nsb->nsb_namespace = ns; nsb->nsb_reclaim_start = 0; @@ -1026,6 +1012,8 @@ out_sysfs: ldlm_namespace_sysfs_unregister(ns); ldlm_namespace_cleanup(ns, 0); out_hash: + OBD_FREE_LARGE(ns->ns_rs_buckets, + BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0])); kfree(ns->ns_name); cfs_hash_putref(ns->ns_rs_hash); out_ns: @@ -1295,6 +1283,8 @@ void ldlm_namespace_free_post(struct ldlm_namespace *ns) ldlm_namespace_debugfs_unregister(ns); ldlm_namespace_sysfs_unregister(ns); cfs_hash_putref(ns->ns_rs_hash); + OBD_FREE_LARGE(ns->ns_rs_buckets, + BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0])); kfree(ns->ns_name); /* Namespace \a ns should be not on list at this time, otherwise * this will cause issues related to using freed \a ns in poold @@ -1504,6 +1494,7 @@ ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, struct cfs_hash_bd bd; __u64 version; int ns_refcount = 0; + int hash; LASSERT(ns != NULL); LASSERT(parent == NULL); @@ -1529,7 +1520,8 @@ ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, if (res == NULL) return ERR_PTR(-ENOMEM); - res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); + hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits); + res->lr_ns_bucket = &ns->ns_rs_buckets[hash]; res->lr_name = *name; res->lr_type = type; -- 1.8.3.1