Re: [PATCH] remoteproc: core: Do not put the device when device_add() fails

[Mathieu Poirier <mathieu.poirier@linaro.org>]

On Wed, Nov 04, 2026 at 01:47:41AM +0800, Yuqiang Sun wrote:
> > Unfortunately I do not have the time to read through and digest what Claude has
> > generated in the changelog.
> 
> Sorry for the long content.
> There is a human-written summary about the bug.
> 
> > The documentation for device_add() is clear, and looking at the implementation
> > of device_add(), I understand why put_device() needs to be called here.
> 
> I agree that as documented in device_add, the put_device must be called
> if it fails.
> 
> But in st_slim_rproc_alloc() (drivers/remoteproc/st_slim_rproc.c), when
> rproc_add() returns the error, the driver calls rproc_free() at its err
> label, which is put_device(&rproc->dev). With the put_device() in
> rproc_add(), that reference is dropped twice, and the driver uses the
> rproc after it is freed.
> 
> The same happens in every other caller: xlnx_r5_remoteproc calls
> rproc_free() too, and the other 23 use devm_rproc_alloc(), whose
> devm_rproc_free() does it.
>

This is the kind of information that is useful to find in a changelog.

> That is why I did not change device_add() or one caller, but
> rproc_add().
>

I agree with your assessment.  The problem is that both success and error path
end up calling rproc_free().  I also agree with your (or the AI's) fix, but add
a comment in the failure path of device_add() that explicitly mention that
put_device() will be called in rproc_free() to avoid reducing the reference
count twice.

And please stop using AI to generate your changelog.
 
> If necessary, I can send out the full document and PoC on
> how to trigger the KASAN.
> 
> Thanks,
> Yuqiang
> 
> Mathieu Poirier <mathieu.poirier@linaro.org> 于2026年11月4日周三 00:28写道：
> >
> > On Thu, Oct 29, 2026 at 09:00:37PM +0800, Yuqiang Sun wrote:
> > > When device_add() fails, rproc_add() calls put_device() and then
> > > removes the character device.  The reference it drops is the one
> > > rproc_alloc() took with device_initialize(), which belongs to the
> > > caller: after a failed rproc_add(), every driver frees the rproc,
> > > either with rproc_free() or through devm_rproc_alloc(), and both drop
> > > that reference again.
> > >
> > > So the rproc is freed inside rproc_add(): by the put_device() itself,
> > > or, with CONFIG_REMOTEPROC_CDEV, by the cdev_del() that follows, when
> > > the character device drops its reference on the parent device.  The
> > > caller then uses the freed rproc and puts it a second time.
> > >
> > > With one allocation failure injected into device_add() (failslab
> > > restricted to rproc_add) while st_fdma probes, KASAN reports:
> > >
> > >   BUG: KASAN: slab-use-after-free in slim_clk_disable+0x38/0xf4 [st_slim_rproc]
> > >   Read of size 8 by task insmod/69
> > >   Call trace:
> > >    slim_clk_disable
> > >    st_slim_rproc_alloc
> > >    st_fdma_probe
> > >   Freed by task 69:
> > >    kfree
> > >    rproc_type_release
> > >    device_release
> > >    kobject_put
> > >    cdev_default_release
> > >    kobject_put
> > >    cdev_del
> > >    rproc_char_device_remove
> > >    rproc_add
> > >    st_slim_rproc_alloc
> > >
> > > followed by use-after-free reports in kobject_put() and
> > > refcount_warn_saturate() from the rproc_free() in
> > > st_slim_rproc_alloc().
> > >
> > > Remove the put_device(): the caller's rproc_free() drops the reference
> > > and releases the rproc, as it does for the other failures in
> > > rproc_add().
> > >
> > > The problem was found with kbound, a static analysis prototype that
> > > checks whether callers rely on an object after a callee returns a value
> > > for which the callee may have released it, on v7.2.12; it is present
> > > in rproc-next.
> > >
> > > Build-tested with x86_64 and arm64 allmodconfig (W=1, sparse): no new
> > > warnings.  Run-tested on QEMU's arm64 virt machine (TCG) with KASAN,
> > > using a device tree node for an ST FDMA with a SLIM core, so that
> > > st_fdma_probe() reaches rproc_add() without the hardware (no clocks,
> > > MMIO windows only mapped), and failing each allocation in rproc_add()
> > > in turn with failslab: before this patch every failure in device_add()
> > > (9 of 9) gives the reports above; with it nothing is reported, and the
> > > rproc is released once, by rproc_free().  Not tested on remoteproc
> > > hardware.
> > >
> > > Fixes: 7dbdb8bd7c02 ("remoteproc: core: Cleanup device in case of failure")
> > > Cc: stable@vger.kernel.org
> > > Assisted-by: Claude:claude-opus-5-5 kbound
> > > Signed-off-by: Yuqiang Sun <izaiahsun@gmail.com>
> > > ---
> > >  drivers/remoteproc/remoteproc_core.c | 4 +---
> > >  1 file changed, 1 insertion(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/remoteproc/remoteproc_core.c b/drivers/remoteproc/remoteproc_core.c
> > > index b88bf179507d..e4f44a3c8b4d 100644
> > > --- a/drivers/remoteproc/remoteproc_core.c
> > > +++ b/drivers/remoteproc/remoteproc_core.c
> > > @@ -2309,10 +2309,8 @@ int rproc_add(struct rproc *rproc)
> > >               return ret;
> > >
> > >       ret = device_add(dev);
> > > -     if (ret < 0) {
> > > -             put_device(dev);
> >
> > The documentation for device_add() is clear, and looking at the implementation
> > of device_add(), I understand why put_device() needs to be called here.
> >
> > Unfortunately I do not have the time to read through and digest what Claude has
> > generated in the changelog.
> >
> > Thanks,
> > Mathieu
> >
> >
> > > +     if (ret < 0)
> > >               goto rproc_remove_cdev;
> > > -     }
> > >
> > >       dev_info(dev, "%s is available\n", rproc->name);
> > >
> > >
> > > base-commit: 82d0bbd67c42a44f0e18b0b7b0da0f3ace0c9527
> > > --
> > > 2.53.0
> > >