Re: [PATCH] fs/locks.c fcntl_setlease did not check if a file was opened for writing before granting a read lease

[Nikita Danilov <Nikita@Namesys.COM>]

Jamie Lokier writes:
 > Joseph D. Wagner wrote:
 > > But I THINK this is how a patch would fix the problem, in theory.
 > 
 > Sorry, it won't.
 > 
 > > +	if ((arg == F_RDLCK)
 > > +	    && ((dentry->d_flags & O_WRONLY)
 > > +		|| (dentry->d_flags & O_RDWR)
 > > +		|| (dentry->d_flags & O_CREAT)
 > > +		|| (dentry->d_flags & O_TRUNC)
 > > +		|| (inode->i_flags & O_WRONLY)
 > > +		|| (inode->i_flags & O_RDWR)
 > > +		|| (inode->i_flags & O_CREAT)
 > > +		|| (inode->i_flags & O_TRUNC)))
 > > +		goto out_unlock;
 > 
 > dentry->d_flags is a combination of the S_ flags, not O_ flags.
 > E.g. S_SYNC, S_NOATIME etc.
 > 
 > inode->i_flags is a combination of the DCACHE_ flags.
 > E.g. DCACHE_AUTOFS_PENDING, DCACHE_REFERENCED tc.

I think it is other way around. ->i_flags are combined from S_SYNC (see,
include/linux/fs.h:IS_IMMUTABLE(), for example), and ->d_flags are
combined from DCACHE_*, latter is explicitly stated in
include/linux/dcache.h

 > 
 > To detect if anyone has the file open for writing, you'll a new count
 > field which keeps track of writer references.  Something like this:
 > 
 > 	if ((arg == F_RDLCK)
 > 	    && ((atomic_read(&inode->i_writer_count) != 0)))
 > 
 > You'll also need to modify all the places where that needs to be
 > maintained.
 > 
 > Btw, I'm not sure why the F_WRLCK case needs to check d_count and
 > i_count.  Isn't it enough to check d_count?  Won't all opens reference
 > the inode through a dentry?:
 > 
 > >  	if ((arg == F_WRLCK)
 > >  	    && ((atomic_read(&dentry->d_count) > 1)
 > >  		|| (atomic_read(&inode->i_count) > 1)))
 > 
 > -- Jamie

Nikita.