Re: [PATCH 0/4] IMA: making i_readcount a first class inode citizen

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, Oct 28, 2010 at 06:02:01PM -0400, Mimi Zohar wrote:
> On Mon, 2010-10-25 at 14:41 -0400, Eric Paris wrote:
> 
> <snip>
> 
> > I believe that IBM is going to look into making i_readcount a first
> > class citizen which can be used by both IMA and generic_setlease().
> > Then people could say IMA had 0 per inode overhead   :)
> 
> This patchset separates the incrementing/decrementing of the i_readcount,
> in the VFS layer, from other IMA functionality, by replacing the current
> ima_counts_get() call with iget_readcount(). Its unclear whether this
> call to increment i_readcount should be made earlier. 
> 
> The patch ordering is a bit redundant in order to leave removing the ifdef
> around i_readcount until the last patch. The first three patches: defines 
> iget/iput_readcount(), moves the IMA functionality in ima_counts_get() to
> ima_file_check(), and removes the IMA imbalance code, simplifying IMA. The
> last patch moves iget/iput_readcount() to the fs directory and removes the
> ifdef around i_readcount, making i_readcount into a "first class inode citizen".
> 
> The generic_setlease code could then take advantage of i_readcount, assuming
> it can take the spin_lock, by doing something like:
> 
> -		if ((arg == F_RDLCK) && (atomic_read(&inode->i_writecount) > 0))
> +
> +		spin_lock(&inode->i_lock);
> +		if ((arg == F_RDLCK) && (atomic_read(&inode->i_writecount) > 0)){
> +			spin_unlock(&inode->i_lock);
>  			goto out;
> -		if ((arg == F_WRLCK)
> -		    && ((atomic_read(&dentry->d_count) > 1)
> -			|| (atomic_read(&inode->i_count) > 1)))
> +		}
> +		if ((arg == F_WRLCK) && (inode->i_readcount > 1)) {
> +			spin_unlock(&inode->i_lock);
>  			goto out;
> +		}
> +		spin_unlock(&inode->i_lock);
>  	}

Seems like an improvement.

It still leaves the race:

	may_open calls lease_break, finds no lease

			setlease checks read/writecount, finds 0,
			creates lease

	__dentry_open bumps read/writecount

(Is there any reason we couldn't move the break_lease to after bumping
read or write count?)

--b.