Re: [PATCH v1] clk: Convert managed get functions to devm_add_action API

[Robin Murphy]

On 2019-12-12 7:10 pm, Dmitry Torokhov wrote:
> On Thu, Dec 12, 2019 at 06:15:16PM +0000, Robin Murphy wrote:
> > On 12/12/2019 4:59 pm, Marc Gonzalez wrote:
> > > On 12/12/2019 15:47, Robin Murphy wrote:
> > >
> > > > On 12/12/2019 1:53 pm, Marc Gonzalez wrote:
> > > >
> > > > > On 11/12/2019 23:28, Dmitry Torokhov wrote:
> > > > >
> > > > > > On Wed, Dec 11, 2019 at 05:17:28PM +0100, Marc Gonzalez wrote:
> > > > > >
> > > > > > > What is the rationale for the devm_add_action API?
> > > > > >
> > > > > > For one-off and maybe complex unwind actions in drivers that wish to use
> > > > > > devm API (as mixing devm and manual release is verboten). Also is often
> > > > > > used when some core subsystem does not provide enough devm APIs.
> > > > >
> > > > > Thanks for the insight, Dmitry. Thanks to Robin too.
> > > > >
> > > > > This is what I understand so far:
> > > > >
> > > > > devm_add_action() is nice because it hides/factorizes the complexity
> > > > > of the devres API, but it incurs a small storage overhead of one
> > > > > pointer per call, which makes it unfit for frequently used actions,
> > > > > such as clk_get.
> > > > >
> > > > > Is that correct?
> > > > >
> > > > > My question is: why not design the API without the small overhead?
> > > >
> > > > Probably because on most architectures, ARCH_KMALLOC_MINALIGN is at
> > > > least as big as two pointers anyway, so this "overhead" should mostly be
> > > > free in practice. Plus the devres API is almost entirely about being
> > > > able to write simple robust code, rather than absolute efficiency - I
> > > > mean, struct devres itself is already 5 pointers large at the absolute
> > > > minimum ;)
> > >
> > > (3 pointers: 1 list_head + 1 function pointer)
> >
> > Ah yes, I failed to mentally preprocess the debug config :)
> >
> > > I'm confused. The first patch was criticized for potentially adding
> > > an extra pointer for every devm_clk_get (e.g. 800 bytes on a 64-bit
> > > platform with 100 clocks).
> >
> > I'm not sure it was a criticism so much as an observation of an aspect that
> > deserved consideration (certainly it was on my part, and I read Dmitry's "It
> > might still, ..." as implying the same). I'd say by this point it has been
> > thoroughly considered, and personally I'm now happy with the conclusion that
> > the kind of embedded platforms that will have many dozens of clocks are also
> > the kind that will tend to have enough padding to make it moot, and thus the
> > code simplification probably is worthwhile overall.
>
> I wonder if we could actually avoid allocating the data with
> ARCH_KMALLOC_MINALIGN in all the cases. It is definitely needed for the
> devm_k*alloc() group of functions as they are direct replacement for
> k*alloc() APIs that give users aligned memory, but for other data
> structures (clocks, regulators, etc, etc) it is not required.

That's a very good point - perhaps something like this (only done properly)?

Robin.

diff --git a/drivers/base/devres.c b/drivers/base/devres.c
index 0bbb328bd17f..2382f963abbe 100644
--- a/drivers/base/devres.c
+++ b/drivers/base/devres.c
@@ -26,14 +26,7 @@ struct devres_node {

 struct devres {
        struct devres_node              node;
-       /*
-        * Some archs want to perform DMA into kmalloc caches
-        * and need a guaranteed alignment larger than
-        * the alignment of a 64-bit integer.
-        * Thus we use ARCH_KMALLOC_MINALIGN here and get exactly the same
-        * buffer alignment as if it was allocated by plain kmalloc().
-        */
-       u8 __aligned(ARCH_KMALLOC_MINALIGN) data[];
+       u8                              data[];
 };

 struct devres_group {
@@ -810,6 +803,17 @@ static int devm_kmalloc_match(struct device *dev, 
void *res, void *data)
 void * devm_kmalloc(struct device *dev, size_t size, gfp_t gfp)
 {
        struct devres *dr;
+       size_t align;
+
+       /*
+        * Some archs want to perform DMA into kmalloc caches
+        * and need a guaranteed alignment larger than
+        * the alignment of a 64-bit integer.
+        * Thus we use ARCH_KMALLOC_MINALIGN here and get exactly the same
+        * buffer alignment as if it was allocated by plain kmalloc().
+        */
+       align = (ARCH_KMALLOC_MINALIGN - sizeof(*dr)) % 
ARCH_KMALLOC_MINALIGN;
+       size += align;

        /* use raw alloc_dr for kmalloc caller tracing */
        dr = alloc_dr(devm_kmalloc_release, size, gfp, dev_to_node(dev));
@@ -822,7 +826,7 @@ void * devm_kmalloc(struct device *dev, size_t size, 
gfp_t gfp)
         */
        set_node_dbginfo(&dr->node, "devm_kzalloc_release", size);
        devres_add(dev, dr->data);
-       return dr->data;
+       return dr->data + align;
 }
 EXPORT_SYMBOL_GPL(devm_kmalloc);