[lustre-devel] [PATCH 03/10] staging: lustre: lu_object: discard extra lru count.

[NeilBrown]

On Tue, May 01 2018, Dilger, Andreas wrote:

> On Apr 30, 2018, at 21:52, NeilBrown <neilb at 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 at 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 at 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
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20180504/7cb2af64/attachment.sig>