On Sat, Jul 14, 2018 at 07:34:45PM +0100, Al Viro wrote:
> On Sat, Jul 14, 2018 at 11:00:32AM -0700, Linus Torvalds wrote:
> > On Sat, Jul 14, 2018 at 10:35 AM Pavel Machek <pavel@xxxxxx> wrote:
> > >
> > > Could we allocate -ve entries from separate slab?
> >
> > No, because negative dentrires don't stay negative.
> >
> > Every single positive dentry starts out as a negative dentry that is
> > passed in to "lookup()" to maybe be made positive.
> >
> > And most of the time they <i>do</i> turn positive, because most of the
> > time people actually open files that exist.
> >
> > But then occasionally you don't, because you're just blindly opening a
> > filename whether it exists or not (to _check_ whether it's there).
>
> BTW, one point that might not be realized by everyone: negative dentries
> are *not* the hard case.
> mount -t tmpfs none /mnt
> touch /mnt/a
> for i in `seq 100000`; do ln /mnt/a /mnt/$i; done
>
> and you've got 100000 *unevictable* dentries, with the time per iteration
> being not all that high (especially if you just call link(2) in a loop).
> They are all positive and all pinned. And you've got only one inode
> there and no persistently opened files, so rlimit and quota won't help
> any.
OK, this
/*
* No ordinary (disk based) filesystem counts links as inodes;
* but each new link needs a new dentry, pinning lowmem, and
* tmpfs dentries cannot be pruned until they are unlinked.
*/
ret = shmem_reserve_inode(inode->i_sb);
if (ret)
goto out;
will probably help (on ramfs it won't, though).
