Re: [BUG] usb: dwc3: Kernel NULL pointer dereference in dwc3_remove()

[Jack Pham]

Hi Sandeep,

On Mon, Jun 14, 2021 at 01:07:11PM +0530, Sandeep Maheswaram wrote:
> 
> On 6/10/2021 9:03 PM, Jack Pham wrote:
> > On Thu, Jun 10, 2021 at 01:11:42PM +0300, Felipe Balbi wrote:
> > > Jack Pham <jackp@xxxxxxxxxxxxxx> writes:
> > > > On Fri, Jun 04, 2021 at 11:20:12AM +0300, Felipe Balbi wrote:
> > > > > Jack Pham <jackp@xxxxxxxxxxxxxx> writes:
> > > > > > > > > > > Alexandru Elisei <alexandru.elisei@xxxxxxx> writes:
> > > > > > > > > > > > I've been able to bisect the panic and the offending commit is 568262bf5492 ("usb:
> > > > > > > > > > > > dwc3: core: Add shutdown callback for dwc3"). I can provide more diagnostic
> > > > > > > > > > > > information if needed and I can help test the fix.
> > > > > > > > > > > if you simply revert that commit in HEAD, does the problem really go
> > > > > > > > > > > away?
> > > > > > > > > > Kernel built from commit 324c92e5e0ee, which is the kernel tip today, the panic is
> > > > > > > > > > there. Reverting the offending commit, 568262bf5492, makes the panic disappear.
> > > > > > > > > Want to send a revert so I can take it now?
> > > > > > > > I can send a revert, but Felipe was asking Sandeep (the commit author) for a fix,
> > > > > > > > so I'll leave it up to Felipe to decide how to proceed.
> > > > > > > I'm okay with a revert. Feel free to add my Acked-by: Felipe Balbi
> > > > > > > <balbi@xxxxxxxxxx> or it.
> > > > > > > 
> > > > > > > Sandeep, please send a new version that doesn't encounter the same
> > > > > > > issue. Make sure to test by reloading the driver in a tight loop for
> > > > > > > several iterations.
> > > > > > This would probably be tricky to test on other "glue" drivers as the
> > > > > > problem appears to be specific only to dwc3_of_simple.  It looks like
> > > > > > both dwc3_of_simple and the dwc3 core now (due to 568262bf5492) each
> > > > > > implement respective .shutdown callbacks. The latter is simply a wrapper
> > > > > > around dwc3_remove(). And from the panic call stack above we see that
> > > > > > dwc3_of_simple_shutdown() calls of_platform_depopulate() which will
> > > > > > again call dwc3_remove() resulting in the double remove.
> > > > > > 
> > > > > > So would an alternative approach be to protect against dwc3_remove()
> > > > > > getting called multiple times? IMO it'd be a bit messy to have to add
> > > > > no, I  don't think so. That sounds like a workaround. We should be able
> > > > > to guarantee that ->remove() doesn't get called twice using the driver
> > > > > model properly.
> > > > Completely fair.  So then having a .shutdown callback that directly calls
> > > > dwc3_remove() is probably not the right thing to do as it completely
> > > > bypasses the driver model so if and when the driver core does later
> > > > release the device from the driver that's how we end up with the double
> > > > remove.
> > > yeah, I would agree with that.
> > > 
> > > > > > additional checks there to know if it had already been called. So maybe
> > > > > > avoid it altogether--should dwc3_of_simple_shutdown() just skip calling
> > > > > > of_platform_depopulate()?
> > > > > I don't know what the idiomatic is nowadays, but at least early on, we
> > > > > had to call depopulate.
> > > > So any suggestions on how to fix the original issue Sandeep was trying
> > > > to fix with 568262bf5492? Maybe implement .shutdown in dwc3_qcom and have
> > > > it follow what dwc3_of_simple does with of_platform_depopulate()? But
> > > > then wouldn't other "glues" want/need to follow suit?
> > > I think we can implement shutdown in core, but we need to careful with
> > > it. Instead of just blindly calling remove, let's extract the common
> > > parts to another internal function that both remove and shutdown
> > > call. debugfs removal should not be part of that generic method :-)
> > Hi Sandeep,
> > 
> > Upon re-reading your description in 568262bf5492 it sounds like the
> > original intention of your patch is basically to quiesce the HW so that
> > it doesn't continue to run after SMMU/IOMMU is disabled right?
> > 
> > If that is the case, couldn't we simply call only dwc3_core_exit_mode()
> > assuming there is no other requirement to do any other cleanup/teardown
> > (PHYs, clocks, resets, runtime PM, et al)? This function should do the
> > bare minimum of stopping the controller in whatever mode (host or
> > peripheral) it is currently operating in.
> 
> Yes that was the intention. I will call only dwc3_core_exit_mode()
> and check. Is there any way we can do from dwc3 qcom glue driver to
> avoid problems for other glue drivers?

As I mentioned above maybe you could just implement a dwc3_qcom specific
.shutdown callback which mimics what dwc3_of_simple() does by calling
of_platform_depopulate(). This will allow the kernel driver core to
invoke dwc3_remove() rather than calling it directly yourself.

The downside is that if other glue drivers want to follow this they'd
have to duplicate the same logic. But maybe this is a more cautious
approach until we start seeing other drivers needing this generically
within core.c.

> > > Anything in that generic method should, probably, be idempotent.
> > Yes we'll need to ensure that dwc3_core_exit_mode() can be called
> > multiple times without additional side effects. At first glance this
> > probably means setting dwc->xhci and dwc->gadget to NULL from
> > dwc3_host_exit() and dwc3_gadget_exit(), respectively.
> 
> Ok. Is there any way to test this ?

You could implement both the dwc3_qcom_shutdown() as above as well as
adding back dwc3_shutdown() which only does dwc3_core_exit_mode(). Make
sure that even though dwc3_core_exit_mode() gets called twice nothing
bad happens.

Jack
-- 
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project