Re: [PATCH 1/2] regmap: teach regmap to use raw spinlocks if requested in the config

[Rasmus Villemoes <linux@rasmusvillemoes.dk>]

On 27/08/2021 01.01, Thomas Gleixner wrote:
> On Wed, Aug 25 2021 at 23:50, Vladimir Oltean wrote:
> 
>> Some drivers might access regmap in a context where a raw spinlock is
>> held. An example is drivers/irqchip/irq-ls-extirq.c, which calls
>> regmap_update_bits() from struct irq_chip :: irq_set_type, which is a
>> method called by __irq_set_trigger() under the desc->lock raw spin lock.
>>
>> Since desc->lock is a raw spin lock and the regmap internal lock for
>> mmio is a plain spinlock (which can become sleepable on RT), this is an
>> invalid locking scheme and we get a splat stating that this is a
>> "[ BUG: Invalid wait context ]".
>>
>> It seems reasonable for regmap to have an option use a raw spinlock too,
>> so add that in the config such that drivers can request it.
> 
> What's reasonable about that?
> 
> What exactly prevents the regmap locking to use a raw spinlock
> unconditionally?

Perhaps this:

        /*
         * When we write in fast-paths with regmap_bulk_write() don't
allocate
         * scratch buffers with sleeping allocations.
         */
        if ((bus && bus->fast_io) || config->fast_io)
                map->alloc_flags = GFP_ATOMIC;
        else
                map->alloc_flags = GFP_KERNEL;

i.e. the regmap code can actually do allocations under whatever internal
lock it uses. So ISTM that any regmap that uses a raw_spinlock (whether
unconditionally or via Vladimir's opt-in) cannot be used with
regmap_bulk_write().

ISTM using regmap for mmio makes things more complicated than necessary.

Rasmus