[RFC PATCH 1/5] of: introduce the overlay manager

[Mathieu Poirier]

On 27 October 2016 at 08:03, Antoine Tenart
<antoine.tenart at free-electrons.com> wrote:
> Hello Mathieu,
>
> On Wed, Oct 26, 2016 at 10:29:59AM -0600, Mathieu Poirier wrote:
>>
>> Please find my comments below.
>
> Thanks for the comments. I expected more distant reviews, on the overall
> architecture to know if this could fit the needs of others. But anyway
> your comments are helpful if we ever decide to go with an overlay
> manager like this one.

I agree - something like this should have attracted more reviews.  I
suggest providing a better explanation, i.e what you are doing and why
in more details.  I reviewed your set and I'm still not sure of
exactly what it does, hence not being able to comment on the validity
of the approach.

I'm pretty sure there is value in what you are suggesting, it's a
matter of getting people to understand the approach and motivation.

>
>> On 26 October 2016 at 08:57, Antoine Tenart
>> <antoine.tenart at free-electrons.com> wrote:
>> > +
>> > +/*
>> > + * overlay_mgr_register_format()
>> > + *
>> > + * Adds a new format candidate to the list of supported formats. The registered
>> > + * formats are used to parse the headers stored on the dips.
>> > + */
>> > +int overlay_mgr_register_format(struct overlay_mgr_format *candidate)
>> > +{
>> > +       struct overlay_mgr_format *format;
>> > +       int err = 0;
>> > +
>> > +       spin_lock(&overlay_mgr_format_lock);
>> > +
>> > +       /* Check if the format is already registered */
>> > +       list_for_each_entry(format, &overlay_mgr_formats, list) {
>> > +               if (!strcpy(format->name, candidate->name)) {
>>
>> This function is public to the rest of the kernel - limiting the
>> lenght of ->name and using strncpy() is probably a good idea.
>
> I totally agree. This was in fact something I wanted to do.
>
>> > +
>> > +/*
>> > + * overlay_mgr_parse()
>> > + *
>> > + * Parse raw data with registered format parsers. Fills the candidate string if
>> > + * one parser understood the raw data format.
>> > + */
>> > +int overlay_mgr_parse(struct device *dev, void *data, char ***candidates,
>>
>> I'm pretty sure there is another way to proceed than using 3 levels of
>> references.  It makes the code hard to read and a prime candidate for
>> errors.
>
> Sure. I guess we could allocate an array of fixed-length strings which
> would be less flexible but I don't think we need something that flexible
> here.
>
>>
>> > +
>> > +               format->parse(dev, data, candidates, n);
>>
>> ->parse() returns an error code.  It is probably a good idea to check
>> it.  If it isn't needed then a comment explaining why it is the case
>> would be appreciated.
>
> So the point of the parse function is to determine if the data read from
> a source is a compatible header of a given format. Returning an error
> doesn't mean the header won't be recognized by another one.
>
> We could maybe handle this better, by returning an error iif different
> that -EINVAL. Or we could have one function to check the compatibility
> and one to parse it, if compatible.
>
>> > +static int _overlay_mgr_apply(struct device *dev, char *candidate)
>> > +{
>> > +       struct overlay_mgr_overlay *overlay;
>> > +       struct device_node *node;
>> > +       const struct firmware *firmware;
>> > +       char *firmware_name;
>> > +       int err = 0;
>> > +
>> > +       spin_lock(&overlay_mgr_lock);
>> > +
>> > +       list_for_each_entry(overlay, &overlay_mgr_overlays, list) {
>> > +               if (!strcmp(overlay->name, candidate)) {
>> > +                       dev_err(dev, "overlay already loaded\n");
>> > +                       err = -EEXIST;
>> > +                       goto err_lock;
>> > +               }
>> > +       }
>> > +
>> > +       overlay = devm_kzalloc(dev, sizeof(*overlay), GFP_KERNEL);
>>
>> Function devm_kzalloc() can sleep but you're holding a spinlock - I'm
>> surprised the kernel didn't complain here.  Allocate the memory before
>> holding the lock.  If the overly is already loaded simply free it on
>> the error path.
>
> Right.
>
> Thanks,
>
> Antoine
>
> --
> Antoine Ténart, Free Electrons
> Embedded Linux and Kernel engineering
> http://free-electrons.com