Re: [PATCH] binder: Address corner cases in deferred copy and fixup

[Alessandro Astone]

On 13/04/2022 12:00, Greg KH wrote:

> On Wed, Apr 13, 2022 at 10:54:27AM +0200, Alessandro Astone wrote:
> > When handling BINDER_TYPE_FDA object we are pushing a parent fixup
> > with a certain skip_size but no scatter-gather copy object, since
> > the copy is handled standalone.
> > If BINDER_TYPE_FDA is the last children the scatter-gather copy
> > loop will never stop to skip it, thus we are left with an item in
> > the parent fixup list. This will trigger the BUG_ON().
> >
> > Furthermore, it is possible to receive BINDER_TYPE_FDA object
> > with num_fds=0 which will confuse the scatter-gather code.
> >
> > In the android userspace I could only find these usecases in the
> > libstagefright OMX implementation, so it might be that they're
> > doing something very weird, but nonetheless the kernel should not
> > panic about it.
> >
> > Fixes: 09184ae9b575 ("binder: defer copies of pre-patched txn data")
> > Signed-off-by: Alessandro Astone <ales.astone@xxxxxxxxx>
> > ---
> >   drivers/android/binder.c | 11 +++++++++--
> >   1 file changed, 9 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/android/binder.c b/drivers/android/binder.c
> > index 8351c5638880..18ad6825ba30 100644
> > --- a/drivers/android/binder.c
> > +++ b/drivers/android/binder.c
> > @@ -2295,7 +2295,7 @@ static int binder_do_deferred_txn_copies(struct binder_alloc *alloc,
> >   {
> >   	int ret = 0;
> >   	struct binder_sg_copy *sgc, *tmpsgc;
> > -	struct binder_ptr_fixup *pf =
> > +	struct binder_ptr_fixup *tmppf, *pf =
> Just make this a new line:
> 	struct binder_ptr_fixup *tmppf;
> above the existing line.
Ack.

> >   		list_first_entry_or_null(pf_head, struct binder_ptr_fixup,
> >   					 node);
> >   
> > @@ -2349,7 +2349,11 @@ static int binder_do_deferred_txn_copies(struct binder_alloc *alloc,
> >   		list_del(&sgc->node);
> >   		kfree(sgc);
> >   	}
> > -	BUG_ON(!list_empty(pf_head));
> So you are hitting this BUG_ON() today?
Correct, both on 5.17, stable 5.17.2 and current master

> > +	list_for_each_entry_safe(pf, tmppf, pf_head, node) {
> > +		BUG_ON(pf->skip_size == 0);
> > +		list_del(&pf->node);
> > +		kfree(pf);
> > +	}
> >   	BUG_ON(!list_empty(sgc_head));
> >   
> >   	return ret > 0 ? -EINVAL : ret;
> > @@ -2486,6 +2490,9 @@ static int binder_translate_fd_array(struct list_head *pf_head,
> >   	struct binder_proc *proc = thread->proc;
> >   	int ret;
> >   
> > +	if (fda->num_fds == 0)
> > +		return 0;
> Why return 0?
>
> This feels like a separate issue from above, should this be 2 different
> commits?
return 0 because I want it to be handled as it was handled before
	09184ae9b575 ("binder: defer copies of pre-patched txn data")

Function `binder_do_deferred_txn_copies` distinguishes between a copy-fixup
and a skip with `if (pf->skip_size)` so if the skip_size is 0, which happens
if fda->num_fds is 0, it would accidentally enter the wrong branch.
By returning 0 early i make sure a skip of size 0 is not added. It's not an
error because it was never an error before commit 09184ae9b575 and some
userspace in android is hitting this path.

I would agree it's a separate issue.
I originally merged it in this same patch because
1) Both are fixups to 09184ae9b575
2) Both are triggered by the same real-world android transaction that looks
   something like this:
     obj[0] BINDER_TYPE_PTR, parent
     obj[1] BINDER_TYPE_PTR, child
     obj[2] BINDER_TYPE_PTR, child
     obj[3] BINDER_TYPE_FDA with num_fds=0
3) In the other hunk of this patch i replace the BUG_ON with:
     BUG_ON(pf->skip_size == 0)
   to only BUG if any item remaining in the pf_head list is not a skip,
   but as observed we may receive skips of size 0.
4) With this hunk only, you would no longer reproduce the BUG_ON because the
   only transaction we receive in android with BINDER_TYPE_FDA as last child
   coincidentally always has num_fds=0. Certainly some weird behaviour...

So if I split them, patch A would depend on patch B (see point 3), but the
BUG of patch B would only be reproducible without patch A (see point 4).

But let me know if you still prefer them split.

> thanks,
>
> greg k-h