On Tue, Jan 14, 2020 at 05:40:34PM +0200, Amir Goldstein wrote:
> Generalize the take_name_snapshot()/release_name_snapshot() interface
> so it is possible to snapshot either a dentry d_name or its snapshot.
>
> The order of fields d_name and d_inode in struct dentry is swapped
> so d_name is adjacent to d_iname. This does not change struct size
> nor cache lines alignment.
>
> Currently, we snapshot the old name in vfs_rename() and we snapshot the
> snapshot the dentry name in __fsnotify_parent() and then we pass qstr
> to inotify which allocated a variable length event struct and copied the
> name.
>
> This new interface allows us to snapshot the name directly into an
> fanotify event struct instead of allocating a variable length struct
> and copying the name to it.
Ugh... That looks like being too damn cute for no good reason. That
trick with union is _probably_ safe, but I wouldn't bet a dime on
e.g. randomize_layout crap not screwing it over in random subset of
gcc versions. You are relying upon fairly subtle reading of 6.2.7
and it feels like just the place for layout-mangling plugins to fuck
up.
With -fplan9-extensions we could go for renaming struct name_snapshot
fields and using an anon member in struct dentry -
...
struct inode *d_inode;
struct name_snapshot; // d_name and d_iname
struct lockref d_lockref;
...
but IMO it's much safer to have an explicit
// NOTE: release_dentry_name_snapshot() will be needed for both copies.
clone_name_snapshot(struct name_snapshot *to, const struct name_snapshot *from)
{
to->name = from->name;
if (likely(to->name.name == from->inline_name)) {
memcpy(to->inline_name, from->inline_name,
to->name.len);
to->name.name = to->inline_name;
} else {
struct external_name *p;
p = container_of(to->name.name, struct external_name, name[0]);
atomic_inc(&p->u.count);
}
}
and be done with that. Avoids any extensions or tree-wide renamings, etc.
