Message-ID: <alpine.LFD.2.00.1107081643510.5442@dhcp-27-109.brq.redhat.com>
Date:	Fri, 8 Jul 2011 16:44:48 +0200 (CEST)
From:	Lukas Czerner <lczerner@...hat.com>
To:	Kyungmin Park <kmpark@...radead.org>
cc:	Lukas Czerner <lczerner@...hat.com>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	Linux Filesystem Mailing List <linux-fsdevel@...r.kernel.org>
Subject: Re: ext4: why init the unused block group at batched discard?

On Fri, 8 Jul 2011, Kyungmin Park wrote:

> On Thu, Jul 7, 2011 at 11:41 PM, Lukas Czerner <lczerner@...hat.com> wrote:
> > On Thu, 7 Jul 2011, Kyungmin Park wrote:
> >
> >> On Thu, Jul 7, 2011 at 6:24 PM, Lukas Czerner <lczerner@...hat.com> wrote:
> >> > On Wed, 6 Jul 2011, Kyungmin Park wrote:
> >> >
> >> >> Hi Lukas,
> >> >>
> >> >> During code review batched discard support at ext4. I wonder why do
> >> >> you init the uninitialized block group during batched discard.
> >> >> As you know uninitialized block group mean that there's no operation
> >> >> at these blocks.
> >> >> So no need to trim it at all.
> >> >
> >> > What you're describing is another flag, namely EXT4_BG_BLOCK_UNINIT,
> >> > which tells us that there was no allocation from that block bitmap since
> >> > the mkfs (as Amir already pointed out). Flag
> >> > EXT4_GROUP_INFO_NEED_INIT_BIT simply states that there is no buddy
> >> > initialized for this group.
> >> >
> >> > That said the code is perfectly fine, and it should not affect even the
> >> > e2fsck which uses EXT4_BG_BLOCK_UNINIT to skip not used block groups
> >> > since we only change it on allocations.
> >> >
> >> > It is true that after the commit
> >> > 78944086663e6c1b03f3d60bf7610128149be5fc ext4: only load buddy bitmap in
> >> > ext4_trim_fs() when it is needed
> >> > we do not longer need to initialize the buddy right away, but wait ontil
> >> > it is really needed. Actually we do not need it at all, because is when
> >> > we are going to load the buddy the ext4_mb_load_buddy() will check for
> >> > the EXT4_GROUP_INFO_NEED_INIT_BIT and will initialize the buddy for us.
> >> >
> >> > Yongqiang pointed out that we might use EXT4_BG_BLOCK_UNINIT to skip
> >> > group as well, but I do not think that it is a good idea, since the
> >> > initial discard at mkfs time might not be done (we just do not know it),
> >> > so any assumption like this are not right. Moreover there are patches
> >> > from Tao Ma which adds the code for skipping groups which has not been
> >> > freed from since the last fitrim call. Search the list for [PATCH 0/4
> >> > RESEND]  ext4 trim bug fixes and improvement.
> >>
> >> Thank you for all kind explanations.
> >>
> >> Another consideration is that even though batched discard has little
> >> overhead it's not good idea trim it all unused blocks at one time.
> >> since disk used blocks doesn't increased in normal case.
> >> So how about to remember the last allocated block group and trim it
> >> until this block group?
> >
> > Yes, this problem is addressed by patches from Tao as mentioned above,
> > however it is still waiting for merge.
> >
> >>
> >> To reduce the trim time. I also consider the divide the block groups
> >> as several trim area e.g., 1 GiB and trim it sequentially.
> >
> > I am not sure what do you mean. In ext4 allocation groups has a lot
> > smaller sizes (128M for 4k block size) than 1G. Also you can specify
> > that you do want to discard just a part of the filesystem, but you
> > probably did notice that in review, right ?
> 
> My concept is that
> 
> Divide the trim area, from 0 to 1GiB, 1GiB to 2GiB, and so on.

Yes, that is what you can do now.

Regards!
-Lukas

Powered by blists - more mailing lists