Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210504163550.1486337-1-leah.rumancik@gmail.com>
 <20210504163550.1486337-2-leah.rumancik@gmail.com> <20210505212711.GA8532@magnolia>
 <20210505220844.GD8532@magnolia> <20210506155853.GF8532@magnolia> <YJQy/DfuCFyZdwe7@google.com>
In-Reply-To: <YJQy/DfuCFyZdwe7@google.com>
From:   harshad shirwadkar <harshadshirwadkar@gmail.com>
Date:   Fri, 7 May 2021 09:22:24 -0700
Message-ID: <CAD+ocbwK9v=ky+bBtMuVMP+zDNYnTSO2DjPZ4sE1AnYw8iEmew@mail.gmail.com>
Subject: Re: [PATCH v3 2/3] ext4: add ioctl EXT4_IOC_CHECKPOINT
To:     Leah Rumancik <leah.rumancik@gmail.com>
Cc:     "Darrick J. Wong" <djwong@kernel.org>,
        Ext4 Developers List <linux-ext4@vger.kernel.org>,
        "Theodore Y. Ts'o" <tytso@mit.edu>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

> > > > > +               err = jbd2_journal_flush(EXT4_SB(sb)->s_journal,
> > >
> > > Huh.  So we don't flush the filesystem at all, just the journal?  I
> > > don't see anything in the documentation saying that syncfs() is a
> > > prerequisite.
>
> This is just for the journal, good point, I'll update the documentation.
It just occurred to me this morning that we need to ensure that a
*full* commit happens before IOC_CHECKPOINT and not a *fast* commit.
Fast commits cannot be checkpointed, they rely on full commits for
checkpoint operation. So, if a syncfs call results in a fast commit,
the following sequence of events will happen:

* Ext4 writes fast commit information in fast commit area

* When user calls EXT4_IOC_CHECKPOINT, the checkpoint operation would
result in checkpointing everything in the main journal, except things
written in fast commit area

* During the discard phase of EXT4_IOC_CHECKPOINT, fast commit area
will be discarded and thus we'll lose the log updates present in the
fast commit area

However, this isn't a problem today. Syncfs doesn't result in a fast
commit but results in a full commit. But, that can change at some
point in the future. So, unless we can either come up with syncfs()
variant that can induce a full commit (which would be a little ugly -
I don't think the user needs to know what kind of journal commit file
system is doing) or add checkpointing support in fast commits, we
should just do a full commit from the IOCTL code.

- Harshad