Re: [PATCH] regmap: move readable check before accessing regcache.

[Srinivas Kandagatla <srinivas.kandagatla@linaro.org>]

Thanks Mark for review,

On 18/06/2021 12:51, Mark Brown wrote:
> On Fri, Jun 18, 2021 at 12:35:58PM +0100, Srinivas Kandagatla wrote:
> 
>> The issue that I encountered is when doing regmap_update_bits on
>> a write only register. In regcache path this will not do the right
>> thing as the register is not readable and driver which is using
>> regmap_update_bits will never notice that it can not do a update
>> bits on write only register leading to inconsistent writes and
>> random hardware behavior.
> 
> Why will use of regmap_update_bits() mean that a driver will never
> notice a write failure?  Shouldn't remgap_update_bits() be fixed to
> report any errors it isn't reporting, or the driver fixed to check

usecase is performing regmap_update_bits() on a *write-only* registers.

_regmap_update_bits() checks _regmap_read() return value before bailing 
out. In non cache path we have this regmap_readable() check however in 
cached patch we do not have this check, so _regmap_read() will return 
success in this case so regmap_update_bits() never reports any error.

driver in question does check the return value.

> error codes?  I really don't understand the issue you're trying to
> report - what is "the right thing" and what makes you believe that a
> driver can't do an _update_bits() on a write only but cached register?
> Can you specify in concrete terms what the problem is.

So one of recent patch ("ASoC: qcom: Fix for DMA interrupt clear reg 
overwriting) 
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commit/?h=next-20210618&id=da0363f7bfd3c32f8d5918e40bfddb9905c86ee1

broke audio on DragonBoard 410c.

This patch simply converts writes to regmap_update_bits for that 
particular dma channel. The register that its updating is IRQ_CLEAR 
register which is software "WRITE-ONLY" and Hardware read-only register.

The bits in particular case is updating is a period interrupt clear bit.

Because we are using regmap cache in this driver,

first regmap_update_bits(map, 0x1, 0x1) on first period interrupt will 
update the cache and write to IRQ_CLEAR hardware register which then 
clears the interrupt latch.
On second period interrupt we do regmap_update_bits(map, 0x1, 0x1) with 
the same bits, Because we are using cache for this regmap caches sees no 
change in the cache value vs the new value so it will never write/update 
  IRQ_CLEAR hardware register, so hardware is stuck here waiting for 
IRQ_CLEAR write from driver and audio keeps repeating the last period.

> 
>> There seems to be missing checks in regcache_read() which is
>> now added by moving the orignal check in _regmap_read() before
>> accessing regcache.
> 
>> Cc: stable@vger.kernel.org
>> Fixes: 5d1729e7f02f ("regmap: Incorporate the regcache core into regmap")
> 
> Are you *sure* you've identified the actual issue here - nobody has seen

I think so, my above triage does summarizes the problem in detail.

> any problems with this in the past decade?  Please don't just pick a
> random commit for the sake of adding a Fixes tag.

I did git blame and picked up this changeset which is when the cache was 
integrated.

> 
>> @@ -2677,6 +2677,9 @@ static int _regmap_read(struct regmap *map, unsigned int reg,
>>   	int ret;
>>   	void *context = _regmap_map_get_context(map);
>>   
>> +	if (!regmap_readable(map, reg))
>> +		return -EIO;
>> +
>>   	if (!map->cache_bypass) {
>>   		ret = regcache_read(map, reg, val);
>>   		if (ret == 0)
>> @@ -2686,9 +2689,6 @@ static int _regmap_read(struct regmap *map, unsigned int reg,
>>   	if (map->cache_only)
>>   		return -EBUSY;
>>   
>> -	if (!regmap_readable(map, reg))
>> -		return -EIO;
>> -
> 
> This puts the readability check before the cache check which will break
> all drivers using the cache on write only registers.
Initially I added check in regcache_read(), later I moved it to 
_regmap_read. do you think check in regcache_read() is the correct place?

--srini
>