Re: [RFC PATCH v2] clk: Use a new helper in managed functions

[Geert Uytterhoeven <geert@linux-m68k.org>]

Hi Marc,

On Thu, Jan 23, 2020 at 1:18 PM Marc Gonzalez <marc.w.gonzalez@free.fr> wrote:
> On 23/01/2020 11:32, Geert Uytterhoeven wrote:
> > On Thu, Jan 23, 2020 at 11:13 AM Marc Gonzalez wrote:
> >> A limitation of devm_add_action is that it stores the void *data argument "as is".
> >> Users cannot pass the address of a struct on the stack. devm_add() addresses that
> >> specific use-case, while being a minimal wrapper around devres_alloc + devres_add.
> >> (devm_add_action adds an extra level of indirection.)
> >
> > I didn't mean the advantage of devm_add() over devm_add_action(),
> > but the advantage of dr_release_t, which has a device pointer.
>
> I'm confused...
>
>         void *devres_alloc(dr_release_t release, size_t size, gfp_t gfp);
>         int devm_add_action(struct device *dev, void (*action)(void *), void *data);
>
> devres_alloc() expects a dr_release_t argument; devm_add() is a thin wrapper
> around devres_alloc(); ergo devm_add() expects that dr_release_t argument.

OK.

> devm_add_action() is a "heavier" wrapper around devres_alloc() which defines
> a "private" release function which calls a user-defined "action".
> (i.e. the extra level of indirection I mentioned above.)
>
> I don't understand the question about the advantage of dr_release_t.

OK. So devm_add_action() is the odd man out there.

> >>>> +       void *data = devres_alloc(func, size, GFP_KERNEL);
> >>>> +
> >>>> +       if (data) {
> >>>> +               memcpy(data, arg, size);
> >>>> +               devres_add(dev, data);
> >>>> +       } else
> >>>> +               func(dev, arg);
> >>>> +
> >>>> +       return data;
> >>>
> >>> Why return data or NULL, instead of 0 or -Efoo, like devm_add_action()?
> >>
> >> My intent is to make devm_add a minimal wrapper (it even started out as
> >> a macro). As such, I just transparently pass the result of devres_alloc.
> >>
> >> Do you see an advantage in processing the result?
> >
> > There are actually two questions to consider here:
> >   1. Is there a use case for returning the data pointer?
> >      I.e. will the caller ever use it?
> >   2. Can there be another failure mode than out-of-memory?
> >      Changing from NULL to ERR_PTR() later means that all callers
> >      need to be updated.
>
> I think I see your point. You're saying it's not good to kick the can down
> the road, because callers won't know what to do with the pointer.

Exactly.

> Actually, I'm in the same boat as these users. I looked at
> devres_alloc -> devres_alloc_node -> alloc_dr -> kmalloc_node_track_caller -> __do_kmalloc
>
> Basically, the result is NULL when something went wrong, but the actual
> error condition is not propagated. It could be:
> 1) check_add_overflow() finds an overflow
> 2) size > KMALLOC_MAX_CACHE_SIZE
> 3) kmalloc_slab() or kasan_kmalloc() fail
> 4) different errors on the CONFIG_NUMA path
>
> Basically, if lower-level functions don't propagate errors, it's not
> easy for a wrapper to do something sensible... ENOMEM looks reasonable
> for kmalloc-related failures.

Indeed.  If devm_add() would return an error code, callers could just check
for error, and propagate the error code, without a need for hardcoding -ENOMEM.

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel