Re: [PATCH 1/1] regmap: regcache-rbtree: Use GFP_ATOMIC when using spinlocks

[Mark Brown <broonie@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 1670 bytes --]

On Mon, Jul 20, 2015 at 09:04:05AM +0200, Alexander Stein wrote:
> On Thursday 16 July 2015 20:06:57, Mark Brown wrote:

Please fix your mailer to word wrap within paragraphs for legibility.

> > The expectation here is that we should either be using no or a flat
> > cache here or (if we're using rbtree) providing register defaults to
> > ensure that we never do allocations in the spinlock.  The rbtree code is
> > written on the assumption that we only have to be faster than reading
> > from a serial bus so I'd be worried about it not behaving at all nicely
> > in a spinlock even ignoring this issue.

> AFAICS even a flat cache seems also only be usefull when providing
> defaults, no? (Or having volatile registers).

Well, it's *better* to provide defaults since otherwise everything
defaults to 0 but it does avoid the whole allocation during fast path
issue since it allocates the cache on init and perhaps that's OK.

> So how to handle this properly? Bail out, if fast_io is available and
> cache_type != (REGCACHE_NONE || REGCACHE_FLAT)?

Or perhaps just if we have to do an allocation?  I can see that someone
might want to use an rbtree and would be careful enough to do the init,
though I *am* a bit dubious about it.

> > Why are you using a dynamically allocated rbtree for a device like this?

> On my way home, I came to the same question. In fact this is not a
> driver written by myself, but from here
> http://git.freescale.com/git/cgit.cgi/ppc/sdk/linux.git/tree/drivers/video/fsl-dcu-fb.c#n1024.
> I guess as this is a mmio device and things like regcache_cache_only()
> are used, REGCACHE_FLAT seems appropriate.

Ah, out of tree BSP code...

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 473 bytes --]