Re: [PATCH] binder: prevent UAF read in print_binder_transaction_log_entry()

[Christian Brauner]

On Tue, Oct 08, 2019 at 02:05:16PM -0400, Joel Fernandes wrote:
> On Tue, Oct 08, 2019 at 03:01:59PM +0200, Christian Brauner wrote:
> > When a binder transaction is initiated on a binder device coming from a
> > binderfs instance, a pointer to the name of the binder device is stashed
> > in the binder_transaction_log_entry's context_name member. Later on it
> > is used to print the name in print_binder_transaction_log_entry(). By
> > the time print_binder_transaction_log_entry() accesses context_name
> > binderfs_evict_inode() might have already freed the associated memory
> > thereby causing a UAF. Do the simple thing and prevent this by copying
> > the name of the binder device instead of stashing a pointer to it.
> > 
> > Reported-by: Jann Horn <jannh@xxxxxxxxxx>
> > Fixes: 03e2e07e3814 ("binder: Make transaction_log available in binderfs")
> > Link: https://lore.kernel.org/r/CAG48ez14Q0-F8LqsvcNbyR2o6gPW8SHXsm4u5jmD9MpsteM2Tw@xxxxxxxxxxxxxx
> > Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
> > Cc: Todd Kjos <tkjos@xxxxxxxxxxx>
> > Cc: Hridya Valsaraju <hridya@xxxxxxxxxx>
> > Signed-off-by: Christian Brauner <christian.brauner@xxxxxxxxxx>
> > ---
> >  drivers/android/binder.c          | 4 +++-
> >  drivers/android/binder_internal.h | 2 +-
> >  2 files changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/android/binder.c b/drivers/android/binder.c
> > index c0a491277aca..5b9ac2122e89 100644
> > --- a/drivers/android/binder.c
> > +++ b/drivers/android/binder.c
> > @@ -57,6 +57,7 @@
> >  #include <linux/sched/signal.h>
> >  #include <linux/sched/mm.h>
> >  #include <linux/seq_file.h>
> > +#include <linux/string.h>
> >  #include <linux/uaccess.h>
> >  #include <linux/pid_namespace.h>
> >  #include <linux/security.h>
> > @@ -66,6 +67,7 @@
> >  #include <linux/task_work.h>
> >  
> >  #include <uapi/linux/android/binder.h>
> > +#include <uapi/linux/android/binderfs.h>
> >  
> >  #include <asm/cacheflush.h>
> >  
> > @@ -2876,7 +2878,7 @@ static void binder_transaction(struct binder_proc *proc,
> >  	e->target_handle = tr->target.handle;
> >  	e->data_size = tr->data_size;
> >  	e->offsets_size = tr->offsets_size;
> > -	e->context_name = proc->context->name;
> > +	strscpy(e->context_name, proc->context->name, BINDERFS_MAX_NAME);
> 
> Strictly speaking, proc-context->name can also be initialized for !BINDERFS
> so the BINDERFS in the MAX_NAME macro is misleading. So probably there should
> be a BINDER_MAX_NAME (and associated checks for whether non BINDERFS names
> fit within the MAX.

I know but I don't think it's worth special-casing non-binderfs devices.
First, non-binderfs devices can only be created through a KCONFIG option
determined at compile time. For stock Android the names are the same for
all vendors afaik.
Second, BINDERFS_MAX_NAME is set to the maximum path name component
length that nearly all filesystems support (256 chars). If you exceed
that then you run afoul of a bunch of other assumptions already and will
cause trouble.
Third, even if there is someone crazy and uses more than 256 chars for a
non-binderfs device at KCONFIG time strscpy will do the right thing and
truncate and you'd see a truncated binder device name. This doesn't seem
to be a big deal for a debugfs interface.
Fourth, the check for non-binderfs devices technically has nothing to do
with this patch. This patch should really just do the minimal thing and
fix the UAF. Which it does.
Fifth, I already tried to push for validation of non-binderfs binder
devices a while back when I wrote binderfs and was told that it's not
needed. Hrydia tried the same and we decided the same thing. So you get
to be the next person to send a patch. :)

> 
> >  	if (reply) {
> 
> >  		binder_inner_proc_lock(proc);
> > diff --git a/drivers/android/binder_internal.h b/drivers/android/binder_internal.h
> > index bd47f7f72075..ae991097d14d 100644
> > --- a/drivers/android/binder_internal.h
> > +++ b/drivers/android/binder_internal.h
> > @@ -130,7 +130,7 @@ struct binder_transaction_log_entry {
> >  	int return_error_line;
> >  	uint32_t return_error;
> >  	uint32_t return_error_param;
> > -	const char *context_name;
> > +	char context_name[BINDERFS_MAX_NAME + 1];
> 
> Same comment here, context_name can be used for non-BINDERFS transactions as
> well such as default binder devices.

See above.

> 
> One more thought, this can be made dependent on CONFIG_BINDERFS since regular
> binder devices cannot be unregistered AFAICS and as Jann said, the problem is
> BINDERFS specific. That way we avoid the memcpy for _every_ transaction.
> These can be thundering when Android starts up.

Unless Todd sees this as a real performance problem I'm weary to
introduce additional checking and record a pointer for non-binderfs and
a memcpy() for binderfs devices. :)

> 
> (I secretly wish C strings could be refcounted to avoid exactly this issue,
> that should not be hard to develop but I am not sure if it is worth it for
> this problem :) - For one, it will avoid having to do the strcpy for _every_
> transaction).
> 
> Other than these nits, please add my tag on whichever is the final solution:
> 
> Reviewed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>

Thanks for the review, Joel. :)
Christian