Re: [RFC PATCH v2 1/3] regmap: Add flat_cache_default_is_zero flag for flat cache

[Sander Vanheule <sander@svanheule.net>]

Hi,

On Mon, 2026-01-12 at 09:58 +0530, Sheetal . wrote:
> From: Sheetal <sheetal@nvidia.com>
> 
> Commit e062bdfdd6ad ("regmap: warn users about uninitialized flat
> cache") added a warning for drivers using REGCACHE_FLAT when reading
> registers not present in reg_defaults.
> 
> For hardware where registers have a power-on-reset value of zero
> or drivers that wish to treat zero as a valid cache default, adding
> all such registers to reg_defaults has drawbacks:
> 
> 1. Maintenance burden: Drivers must list every readable register
>    regardless of its reset value.

If you would extend regmap_config to accept a callback for defaults, this burden
could be reduced to near-zero if it was a simple wrapper for the writeable_regs
and/or readable_regs callbacks. I.e. "return 0x0 if readable or writable".


> 2. No functional benefit: Entries like { REG, 0x0 } only set the
>    validity bit; the flat cache value is already zero.

There is a functional benefit: it allows regmap_sync() to do what it is supposed
to do: check if a cached register matches the reset value to omit writing the
cached value to hardware after a reset.


> 3. Code bloat: Large reg_defaults arrays increase driver size.

As I argued before, bloat would be limited if the user could provide tables or
callbacks for defaults.

> 
> Add a flat_cache_default_is_zero flag to struct regmap_config. When
> set, the flat cache marks registers as valid on first read instead
> of warning.
> 
> The valid bit is set on first read rather than marking all registers
> valid at init time for the following reasons:
> - Avoids writes to register holes or unused addresses during sync.
> - Safer for drivers that don't have writeable_reg callback defined.

These are benefits of using a sparse cache, but on closer inspection don't
actually apply here. You can validate this in your kunit test if you also do the
negative check: assert that registers outside of the written range are left
untouched (written == 0).

Because regcache_read() will never return -ENOENT with this option enabled,
regmap_sync() will see all registers as "written". Since there are no defaults
to check against, this will cause regmap_sync() to write to *all* writable
registers. With this patch, you are just reverting to the old behaviour where
regmap_sync() silently overwrites registers you never accessed with 0.


Best,
Sander