From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-fsdevel-owner@vger.kernel.org>
Received: from mx1.redhat.com ([209.132.183.28]:37074 "EHLO mx1.redhat.com"
        rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP
        id S1754013AbdJIOQZ (ORCPT <rfc822;linux-fsdevel@vger.kernel.org>);
        Mon, 9 Oct 2017 10:16:25 -0400
Subject: Re: [PATCH v7 5/6] lib/dlock-list: Enable faster lookup with hashing
To: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>, Jan Kara <jack@suse.com>,
        Jeff Layton <jlayton@poochiereds.net>,
        "J. Bruce Fields" <bfields@fieldses.org>,
        Tejun Heo <tj@kernel.org>,
        Christoph Lameter <cl@linux-foundation.org>,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        Ingo Molnar <mingo@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Andi Kleen <andi@firstfloor.org>,
        Dave Chinner <dchinner@redhat.com>,
        Boqun Feng <boqun.feng@gmail.com>
References: <1507229008-20569-1-git-send-email-longman@redhat.com>
 <1507229008-20569-6-git-send-email-longman@redhat.com>
 <20171009130827.GC5131@linux-80c1.suse>
From: Waiman Long <longman@redhat.com>
Message-ID: <95cbbab7-246d-fd42-23e3-19aa5bd8efe5@redhat.com>
Date: Mon, 9 Oct 2017 10:16:21 -0400
MIME-Version: 1.0
In-Reply-To: <20171009130827.GC5131@linux-80c1.suse>
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 8BIT
Content-Language: en-US
Sender: linux-fsdevel-owner@vger.kernel.org
List-ID: <linux-fsdevel.vger.kernel.org>

On 10/09/2017 09:08 AM, Davidlohr Bueso wrote:
> On Thu, 05 Oct 2017, Waiman Long wrote:
>
>> Insertion and deletion is relatively cheap and mostly contention
>> free for dlock-list. Lookup, on the other hand, can be rather costly
>> because all the lists in a dlock-list will have to be iterated.
>>
>> Currently dlock-list insertion is based on the cpu that the task is
>> running on. So a given object can be inserted into any one of the
>> lists depending on what the current cpu is.
>>
>> This patch provides an alternative way of list selection. The caller
>> can provide a object context which will be hashed to one of the list
>> in a dlock-list. The object can then be added into that particular
>> list. Lookup can be done by iterating elements in the provided list
>> only instead of all the lists in a dlock-list.
>
> Unless I'm misusing the api, I could not find a standard way of
> iterating a _particular_ list head (the one the dlock_list_hash()
> returned). This is because iterators always want the all the heads.
>
> Also, in my particular epoll case I'd need the head->lock _not_ to
> be dropped after the iteration, and therefore is pretty adhoc.
> Currently we do:
>
> dlist_for_each_entry() {
>     // acquire head->lock for each list
> }
> // no locks held
> dlist_add()
>
> I'm thinking perhaps we could have dlist_check_add() which passes a
> callback to ensure we want to add the node. The function could acquire
> the head->lock and not release it until the very end. 

With the dlock_list_hash(), dlock-list is degenerated into a pool of
list where one is chosen by hashing. So the regular list iteration
macros like list_for_each_entry() can then be used. Of course, you have
to explicitly do the lock and unlock operation.

I could also encapsulate it a bit with inlined function like

dlock_list_single_iter_init(iter, dlist, head, flags)

It could set up the iterator properly to iterate only 1 list. The flags
can be to indicate holding the lock after iteration. In this case,
dlock_list_unlock(iter) will have to be called to do the unlock. I could
add a patch to do that if you prefer that route.

Cheers,
Longman