Re: [PATCH v2 1/4] dcache: sweep cached negative dentries to the end of list of siblings

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Feb 16, 2022 at 03:27:39AM +0000, Al Viro wrote:
> On Tue, Feb 15, 2022 at 06:24:53PM -0800, Stephen Brennan wrote:
> 
> > It seems to me that, if we had taken a reference on child by
> > incrementing the reference count prior to unlocking it, then
> > dentry_unlist could never have been called, since we would never have
> > made it into __dentry_kill. child would still be on the list, and any
> > cursor (or sweep_negative) list updates would now be reflected in
> > child->d_child.next. But dput is definitely not safe while holding a
> > lock on a parent dentry (even more so now thanks to my patch), so that
> > is out of the question.
> > 
> > Would dput_to_list be an appropriate solution to that issue? We can
> > maintain a dispose list in d_walk and then for any dput which really
> > drops the refcount to 0, we can handle them after d_walk is done. It
> > shouldn't be that many dentries anyway.
> 
> 	Interesting idea, but... what happens to behaviour of e.g.
> shrink_dcache_parent()?  You'd obviously need to modify the test in
> select_collect(), but then the selected dentries become likely candidates
> for d_walk() itself wanting to move them over to its internal shrink list.
> OTOH, __dput_to_list() will just decrement the count and skip the sucker
> if it's already on a shrink list...
> 
> 	It might work, but it really needs a careful analysis wrt.
> parallel d_walk().  What happens when you have two threads hitting
> shrink_dcache_parent() on two different places, one being an ancestor
> of another?  That can happen in parallel, and currently it does work
> correctly, but that's fairly delicate and there are places where a minor
> change could turn O(n) into O(n^2), etc.
> 
> 	Let me think about that - I'm not saying it's hopeless, and it
> would be nice to avoid that subtlety in dentry_unlist(), but there
> might be dragons.

PS: another obvious change is that d_walk() would become blocking.
So e.g.

int path_has_submounts(const struct path *parent)
{
        struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };

	read_seqlock_excl(&mount_lock);
	d_walk(parent->dentry, &data, path_check_mount);
	read_sequnlock_excl(&mount_lock);

	return data.mounted;
} 

would need a rework - d_walk() is under a spinlock here.  Another
potential headache in that respect is d_genocide() - currently non-blocking,
with this change extremely likely to do evictions.  That, however, is
not a problem for current in-tree callers - they are all shortly followed
by shrink_dcache_parent() or equivalents.

path_has_submounts(), though...  I'd really hate to reintroduce the
"call this on entry/call this on exit" callbacks.  Perhaps it would
be better to pass the dispose list to d_walk() and have the callers
deal with evictions?  For that matter, shrink_dcache_parent() and
friends would be just fine passing the same list they are collecting
into.

<looks at path_has_submounts() callers>
*growl*
autofs_d_automount() has it called under sbi->fs_lock.  So we'd need
to take the disposal all the way out there, and export shrink_dentry_list()
while we are at it.  Not pretty ;-/

And no, we can't make the disposal async, so offloading it to a worker or
thread is not feasible...