Messages in this thread | | | From | NeilBrown <> | Date | Fri, 04 May 2018 10:08:35 +1000 | Subject | Re: [PATCH 03/10] staging: lustre: lu_object: discard extra lru count. |
| |
On Tue, May 01 2018, Dilger, Andreas wrote:
> On Apr 30, 2018, at 21:52, NeilBrown <neilb@suse.com> wrote: >> >> lu_object maintains 2 lru counts. >> One is a per-bucket lsb_lru_len. >> The other is the per-cpu ls_lru_len_counter. >> >> The only times the per-bucket counters are use are: >> - a debug message when an object is added >> - in lu_site_stats_get when all the counters are combined. >> >> The debug message is not essential, and the per-cpu counter >> can be used to get the combined total. >> >> So discard the per-bucket lsb_lru_len. >> >> Signed-off-by: NeilBrown <neilb@suse.com> > > Looks reasonable, though it would also be possible to fix the percpu > functions rather than adding a workaround in this code.
It would be possibly to change percpu_counter_read_positive() to take a const (and we should), but thanks to you highlighting this I realized that I should be using percpu_counter_sum_positive(), to read across all per-cpu counters. That takes a spinlock so it would be a much harder sell to get it to accept a const pointer.
Thanks, NeilBrown
> > Reviewed-by: Andreas Dilger <andreas.dilger@dilger.ca> > >> --- >> drivers/staging/lustre/lustre/obdclass/lu_object.c | 24 ++++++++------------ >> 1 file changed, 9 insertions(+), 15 deletions(-) >> >> diff --git a/drivers/staging/lustre/lustre/obdclass/lu_object.c b/drivers/staging/lustre/lustre/obdclass/lu_object.c >> index 2a8a25d6edb5..2bf089817157 100644 >> --- a/drivers/staging/lustre/lustre/obdclass/lu_object.c >> +++ b/drivers/staging/lustre/lustre/obdclass/lu_object.c >> @@ -57,10 +57,6 @@ >> #include <linux/list.h> >> >> struct lu_site_bkt_data { >> - /** >> - * number of object in this bucket on the lsb_lru list. >> - */ >> - long lsb_lru_len; >> /** >> * LRU list, updated on each access to object. Protected by >> * bucket lock of lu_site::ls_obj_hash. >> @@ -187,10 +183,9 @@ void lu_object_put(const struct lu_env *env, struct lu_object *o) >> if (!lu_object_is_dying(top)) { >> LASSERT(list_empty(&top->loh_lru)); >> list_add_tail(&top->loh_lru, &bkt->lsb_lru); >> - bkt->lsb_lru_len++; >> percpu_counter_inc(&site->ls_lru_len_counter); >> - CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, lru_len: %ld\n", >> - o, site->ls_obj_hash, bkt, bkt->lsb_lru_len); >> + CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p\n", >> + o, site->ls_obj_hash, bkt); >> cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1); >> return; >> } >> @@ -238,7 +233,6 @@ void lu_object_unhash(const struct lu_env *env, struct lu_object *o) >> >> list_del_init(&top->loh_lru); >> bkt = cfs_hash_bd_extra_get(obj_hash, &bd); >> - bkt->lsb_lru_len--; >> percpu_counter_dec(&site->ls_lru_len_counter); >> } >> cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash); >> @@ -422,7 +416,6 @@ int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s, >> cfs_hash_bd_del_locked(s->ls_obj_hash, >> &bd2, &h->loh_hash); >> list_move(&h->loh_lru, &dispose); >> - bkt->lsb_lru_len--; >> percpu_counter_dec(&s->ls_lru_len_counter); >> if (did_sth == 0) >> did_sth = 1; >> @@ -621,7 +614,6 @@ static struct lu_object *htable_lookup(struct lu_site *s, >> lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT); >> if (!list_empty(&h->loh_lru)) { >> list_del_init(&h->loh_lru); >> - bkt->lsb_lru_len--; >> percpu_counter_dec(&s->ls_lru_len_counter); >> } >> return lu_object_top(h); >> @@ -1834,19 +1826,21 @@ struct lu_site_stats { >> unsigned int lss_busy; >> }; >> >> -static void lu_site_stats_get(struct cfs_hash *hs, >> +static void lu_site_stats_get(const struct lu_site *s, >> struct lu_site_stats *stats, int populated) >> { >> + struct cfs_hash *hs = s->ls_obj_hash; >> struct cfs_hash_bd bd; >> unsigned int i; >> + /* percpu_counter_read_positive() won't accept a const pointer */ >> + struct lu_site *s2 = (struct lu_site *)s; > > It would seem worthwhile to change the percpu_counter_read_positive() and > percpu_counter_read() arguments to be "const struct percpu_counter *fbc", > rather than doing this cast here. I can't see any reason that would be bad, > since both implementations just access fbc->count, and do not modify anything. > >> + stats->lss_busy += cfs_hash_size_get(hs) - >> + percpu_counter_read_positive(&s2->ls_lru_len_counter); >> cfs_hash_for_each_bucket(hs, &bd, i) { >> - struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd); >> struct hlist_head *hhead; >> >> cfs_hash_bd_lock(hs, &bd, 1); >> - stats->lss_busy += >> - cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len; >> stats->lss_total += cfs_hash_bd_count_get(&bd); >> stats->lss_max_search = max((int)stats->lss_max_search, >> cfs_hash_bd_depmax_get(&bd)); >> @@ -2039,7 +2033,7 @@ int lu_site_stats_print(const struct lu_site *s, struct seq_file *m) >> struct lu_site_stats stats; >> >> memset(&stats, 0, sizeof(stats)); >> - lu_site_stats_get(s->ls_obj_hash, &stats, 1); >> + lu_site_stats_get(s, &stats, 1); >> >> seq_printf(m, "%d/%d %d/%ld %d %d %d %d %d %d %d\n", >> stats.lss_busy, >> >> > > Cheers, Andreas > -- > Andreas Dilger > Lustre Principal Architect > Intel Corporation [unhandled content-type:application/pgp-signature] | |