On Wed, 26 Jul 2000, Jamie Lokier wrote:
> Tigran Aivazian wrote:
> > Actually, the idea of union-ing things is probably broken anyway. Imagine
> > clear_inode() is called on a pipe (because pipefs does iput(), e.g. from
> > get_pipe_inode()):
> >
> > if (inode->i_bdev) {
> > bdput(inode->i_bdev);
> > inode->i_bdev = NULL;
> > }
> >
> > in the union version, for a pipe, inode->i_bdev may well be not NULL and
> > so bdput() will be called illegally. So, my yesterday's patch appears to
> > be broken. (which is why I left Linus in cc to this message, I would not
> > have wasted his time otherwise).
>
> 1. Make the name explicit: inode->i_union->bdev.
I don't think renaming things make them any different.
>
> 2. Change all the code that refers to the newly unioned fields.
> It should be clear from the name i_union that no code should
> read a field until it knows the type of the inode.
ah, that is much better - what you are saying is the above code should
look like:
if (IS_BLK(inode->i_mode) & inode->i_bdev) {
bdput(inode->i_bdev);
inode->i_bdev = NULL;
}
Then it is a typical tradeoff between speed and memory optimizations - we
gain 4 bytes in the inode but get an extra comparison instruction in the
code. To me, speed is more important - extra memory one can always buy but
the time... hmmm redeem the time for the days are evil.
Regards,
Tigran
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon Jul 31 2000 - 21:00:21 EST