Re: Reviving planned ldso changes

[Rich Felker <dalias@libc.org>]

On Sat, Mar 04, 2017 at 11:58:18AM +0100, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2017-03-02 20:30:26 -0500]:
> > Here's a v4 of the patch that saves the "init parent" we descended
> > from so that it can return where it left off. There are a couple
> > gratuitous hunks left over adding setting of "needed_by" where it made
> > sense to be set, but it's not actually used anymore. They could be
> > dropped if desired but are probably nice to keep for the sake of
> > consistency of data, even thoough it's data we don't use.
> > 
> > I believe this can be extended to allow concurrent dlopen by amending
> > the case in the tree-walk where a dependency isn't constructed yet but
> > already has an "init parent" to check whether it's
> > pending-construction in the calling thread (recursive dlopen from a
> > ctor) or another thread; in the former case (as now) treat it as
> > already-constructed; in the latter, wait on a condvar that gets
> > signaled at the end of each construction, then continue the loop
> > without advancing p. There are probably some subtleties I'm missing,
> > though.
> ....
> >  static void do_init_fini(struct dso *p)
> >  {
> >  	size_t dyn[DYN_CNT];
> > -	int need_locking = libc.threads_minus_1;
> > -	/* Allow recursive calls that arise when a library calls
> > -	 * dlopen from one of its constructors, but block any
> > -	 * other threads until all ctors have finished. */
> > -	if (need_locking) pthread_mutex_lock(&init_fini_lock);
> > -	for (; p; p=p->prev) {
> > -		if (p->constructed) continue;
> > +	pthread_mutex_lock(&init_fini_lock);
> > +	/* Construct in dependency order without any recursive state. */
> > +	while (p && !p->constructed) {
> > +		/* The following loop descends into the first dependency
> > +		 * that is neither alredy constructed nor pending
> > +		 * construction due to circular deps, stopping only
> > +		 * when it reaches a dso with no remaining dependencies
> > +		 * to descend into. */
> > +		while (p->deps && p->deps[p->next_dep]) {
> > +			if (!p->deps[p->next_dep]->constructed &&
> > +			    !p->deps[p->next_dep]->init_parent) {
> > +				p->deps[p->next_dep]->init_parent = p;
> > +				p = p->deps[p->next_dep++];
> 
> i think the root may be visited twice because it
> has no init_parent, which may be problematic with
> the concurrent dlopen (and can cause unexpected
> ctor order: the root node is not constructed last
> if there is a cycle through it)

Ah, the case where the root is an indirect dependency for itself? Yes,
I think you're right in that case. We should be able to avoid it by
setting the initial p->init_parent to head (the application), I think.

> i think only checking init_parent of a dep is
> enough and the root node can have a dummy parent
> that is guaranteed to be not a dependency (ldso?)
> and constructed so it stops the loop.

I think ldso would work too, but in principle it need not be a
dependency of anything if you have a dynamic-linked program that
doesn't use libc (-nostdlib), so it's better to use head, I think.

Also I agree we don't need to check p->constructed now, but once we
unlock during ctor execution, the !init_parent and !constructed cases
need to be treated separately. If it's constructed or pending
construction in the same thread, we can just do p->next_dep++, but if
it has an init_parent but isn't constructed or pending construction in
same thread (recursive) we need to condvar wait and re-check instead,
right?

Rich