Re: regulator: deadlock vs memory reclaim

[Michał Mirosław]

On Mon, Aug 10, 2020 at 06:31:36PM +0100, Mark Brown wrote:
> On Mon, Aug 10, 2020 at 06:09:36PM +0200, Michał Mirosław wrote:
> > On Mon, Aug 10, 2020 at 04:39:28PM +0100, Mark Brown wrote:
> > > On Mon, Aug 10, 2020 at 12:25:37AM +0200, Michał Mirosław wrote:
> > > > regulator_lock_dependent() starts by taking regulator_list_mutex, The
> > > > same mutex covers eg. regulator initialization, including memory allocations
> > > > that happen there. This will deadlock when you have filesystem on eg. eMMC
> > > > (which uses a regulator to control module voltages) and you register
> > > > a new regulator (hotplug a device?) when under memory pressure.
> > > OK, that's very much a corner case, it only applies in the case of
> > > coupled regulators.  The most obvious thing here would be to move the
> > > allocations on registration out of the locked region, we really only
> > > need this in the regulator_find_coupler() call I think.  If the
> > > regulator isn't coupled we don't need to take the lock at all.
> > Currently, regulator_lock_dependent() is called by eg. regulator_enable() and
> > regulator_get_voltage(), so actually any regulator can deadlock this way.
> The initialization cases that are the trigger are only done for coupled
> regulators though AFAICT, otherwise we're not doing allocations with the
> lock held and should be able to progress.

I caught a few lockdep complaints that suggest otherwise, but I'm still
looking into that.

> > > > Basically, we have a BKL for regulator_enable() and we're using ww_mutex
> > > > as a recursive mutex with no deadlock prevention whatsoever. The locks
> > > > also seem to cover way to much (eg. initialization even before making the
> > > > regulator visible to the system).
> > > Could you be more specific about what you're looking at here?  There's
> > > nothing too obvious jumping out from the code here other than the bit
> > > around the coupling allocation, otherwise it looks like we're locking
> > > list walks.
> > When you look at the regulator API (regulator_enable() and friends),
> > then in their implementation we always start by .._lock_dependent(),
> > which takes regulator_list_mutex around its work. This mutex is what
> > makes the code deadlock-prone vs memory allocations. I have a feeling
> > that this lock is a workaround for historical requirements (recursive
> > locking of regulator_dev) that might be no longer needed or is just
> > too defensive programming. Hence my other patches and this inquiry.
> We need to both walk up the tree for operations that involve the parent
> and rope in any peers that are coupled, the idea is basically to avoid
> changes to the coupling while we're trying to figure that out.

This part I understand and this is what ww_mutex should take care of.
But there is another lock that's taken around ww_mutex locking transaction
that causes the trouble. I would expect that the rdev->mutex should be
enough to guard against changes in coupled regulators list, but this
might need a fix in the registration phase to guarantee that (it
probably should do the same ww_mutex locking dance as .._lock_dependent()
does now).

Best Regards,
Michał Mirosław