Re: wip/fsync branch for review

[Colin Walters <walters verbum org>]

On Fri, Apr 11, 2014 at 6:29 PM, Mrunal Patel <mpatel redhat com> wrote:
> > >  https://git.gnome.org/browse/ostree/log/?h=wip/fsync
>
> > >  Review appreciated!
>
> The changes make sense to me. syncfs could be considered as an 
> alternative in the future as discussed in IRC.

Yeah, it's a good point, we should look at syncfs.  Unlike sync() it 
actually has an error return code.  In the normal case though we're not 
writing a *lot* of data - we're just making a hardlink farm.

On the other hand, the downside of syncfs() is it forces a flush of 
everything - say you have a database that's running while you're trying 
to stream an update, forcing a filesystem sync is a bit painful.

On the *other* other hand, one should really reboot soon after upgrade, 
so you'll be taking the database down anyways.

Maybe the right thing is to go belt, suspenders, and superglue; do 
fsync/fdatasync while we write, call syncfs() on the root partition and 
/boot, and finally a global sync() just for good measure.

Thanks for looking at the patches!