Re: [RFC] write_super is for syncing

[Chris Mason]

On Tuesday, March 12, 2002 02:15:40 PM -0800 Andrew Morton <[e-mail]> wrote:

> Chris Mason wrote:
>> 
>> Hi everyone,
>> 
>> The fact that write_super gets called for both syncs and periodic
>> commits causes problems for the journaled filesystems, since we
>> need to trigger commits on write_super to have sync() behave
>> properly.
>> 
>> So, this patch adds a new super operation called commit_super,
>> and extends struct super.s_dirt a little so the filesystem
>> can say: call me on sync() but don't call me from kupdate.
>> 
>> if (s_dirt & S_SUPER_DIRTY) call me from kupdate and on sync
>> if (s_dirt & S_SUPER_DIRTY_COMMIT) call me on sync only.
>> 
> 
> I'm not quite sure why these flags exist?  Would it not be
> sufficient to just call ->write_super() inside kupdate,
> and ->commit_super in fsync_dev()?  (With a ->write_super
> fallback, of course).

fsync_dev(dev != 0) is easy, you can ignore the dirty flag
and call commit_super on the proper device.

But, the loop in sync_supers(dev == 0) is harder, it expects
some flag it can check, and it expects the callback to the FS
will clear that flag.  Adding a new flag seemed like more fun
than redoing the locking and super walk.  I'm curious to hear what 
Al thinks of it though.

-chris