Re: [RFC][PATCH 2/7] UBC: core (structures, API)

[Kirill Korotaev <dev@sw.ru>]

Oleg,

>>+struct user_beancounter *beancounter_findcreate(uid_t uid,
>>+		struct user_beancounter *p, int mask)
>>+{
>>+	struct user_beancounter *new_ub, *ub, *tmpl_ub;
>>+	unsigned long flags;
>>+	struct hlist_head *slot;
>>+	struct hlist_node *pos;
>>+
>>+	if (mask & UB_LOOKUP_SUB) {
>>+		WARN_ON(p == NULL);
>>+		tmpl_ub = &default_subbeancounter;
>>+		slot = &ub_hash[ub_subhash_fun(p, uid)];
>>+	} else {
>>+		WARN_ON(p != NULL);
>>+		tmpl_ub = &default_beancounter;
>>+		slot = &ub_hash[ub_hash_fun(uid)];
>>+	}
>>+	new_ub = NULL;
>>+
>>+retry:
>>+	spin_lock_irqsave(&ub_hash_lock, flags);
>>+	hlist_for_each_entry (ub, pos, slot, hash)
>>+		if (ub->ub_uid == uid && ub->parent == p)
>>+			break;
>>+
>>+	if (pos != NULL) {
>>+		get_beancounter(ub);
>>+		spin_unlock_irqrestore(&ub_hash_lock, flags);
>>+
>>+		if (new_ub != NULL) {
>>+			put_beancounter(new_ub->parent);
> 
> 					^^^^^^^^^^^^^^
> 
> Stupid question: why ->parent can't be NULL ? (without UB_LOOKUP_SUB).
oh, good catch. We removed the check for NULL in put_beancounter() for cleanup,
but didn't notice this place. Thanks!

>>+	if (mask & UB_ALLOC_ATOMIC) {
>>+		new_ub = kmem_cache_alloc(ub_cachep, GFP_ATOMIC);
>>+		if (new_ub == NULL)
>>+			goto out_unlock;
>>+
>>+		memcpy(new_ub, tmpl_ub, sizeof(*new_ub));
>>+		init_beancounter_struct(new_ub, uid);
>>+		if (p)
>>+			new_ub->parent = get_beancounter(p);
>>+		goto out_install;
>>+	}
>>+
>>+	spin_unlock_irqrestore(&ub_hash_lock, flags);
>>+
>>+	new_ub = kmem_cache_alloc(ub_cachep, GFP_KERNEL);
> 
> 
> beancounter_findcreate() is not time critical, yes? Isn't it better
> to kill 'if (mask & UB_ALLOC_ATOMIC) { ... }' and just do
> 
> 	new_ub = kmem_cache_alloc(ub_cachep,
> 		UB_ALLOC_ATOMIC ? GFP_ATOMIC : GFP_KERNEL);
> 
> after spin_unlock(&ub_hash_lock) ?
ok, will do. Thanks for comments!

Kirill