Re: Reviving planned ldso changes

[Rich Felker <dalias@libc.org>]

On Tue, Mar 07, 2017 at 05:02:09PM -0500, Rich Felker wrote:
> On Sun, Mar 05, 2017 at 08:11:59PM -0500, Rich Felker wrote:
> > 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?
> 
> Arg, deep problems I missed. Quoting from IRC:
> 
> <dalias> nsz, uhg, the dep-order draft so far has a big bug
> <dalias> p->deps is not actually deps for p
> <dalias> rather, it's all indirect deps, but only set for a lib that was explicitly dlopen'd
> <dalias> so the new code doesn't actually do dep-order
> <dalias> it just walks a flat list of all (breadth-first, not depth-first) direct and indirect dependencies of p
> <dalias> and descends into each then immediately backs out
> <dalias> because after descending, p->deps is null
> <dalias> i think we should get rid of the old use of p->deps
> <dalias> which is just undoing temp-globalization of libs during load for reloc purposes
> 
> If I first do the work of having a separate global-namespace dso list
> (which is a pending change that will speed up relocations anyway),
> then the old use of p->deps is no longer needed and we can simply
> repurpose it to be direct-deps only.

This is incorrect. p->deps in its current form is used for dlsym
"dependency ordering" symbol resolution, where a breadth-first list of
all direct and indirect dependencies is exactly what you want. So I
don't think it can be eliminated.

I wonder if it suffices to walk the flat p->deps in reverse. I suspect
there are cases where this is wrong when a dependency appears more
than once.

Rich