[RFC 12/22] dm: add a hidden link to efi object

[AKASHI Takahiro]

On Mon, Oct 11, 2021 at 10:09:19AM -0600, Simon Glass wrote:
> Hi Heinrich,
> 
> On Mon, 11 Oct 2021 at 09:31, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
> >
> >
> >
> > On 10/11/21 16:54, Simon Glass wrote:
> > > Hi Takahiro,
> > >
> > > On Mon, 11 Oct 2021 at 00:43, AKASHI Takahiro
> > > <takahiro.akashi at linaro.org> wrote:
> > >>
> > >> Simon,
> > >>
> > >> On Sun, Oct 10, 2021 at 08:14:18AM -0600, Simon Glass wrote:
> > >>> Hi Takahiro,
> > >>>
> > >>> On Thu, 30 Sept 2021 at 23:04, AKASHI Takahiro
> > >>> <takahiro.akashi at linaro.org> wrote:
> > >>>>
> > >>>> This member field in udevice will be used to dereference from udevice
> > >>>> to efi_object (or efi_handle).
> > >>>>
> > >>>> Signed-off-by: AKASHI Takahiro <takahiro.akashi at linaro.org>
> > >>>> ---
> > >>>>   include/dm/device.h | 4 ++++
> > >>>>   1 file changed, 4 insertions(+)
> > >>>
> > >>> I think this should be generalised.
> > >>>
> > >>> Can we add a simple API for attaching things to devices? Something like:
> > >>
> > >> Ok.
> > >>
> > >>
> > >>> config DM_TAG
> > >>>     bool "Support tags attached to devices"
> > >>>
> > >>> enum dm_tag_t {
> > >>>      DM_TAG_EFI = 0,
> > >>>
> > >>>      DM_TAG_COUNT,
> > >>> };
> > >>>
> > >>> ret = dev_tag_set_ptr(dev, DM_TAG_EFI, ptr);
> > >>>
> > >>> void *ptr = dev_tag_get_ptr(dev, DM_TAG_EFI);
> > >>>
> > >>> ulong val = dev_tag_get_val(dev, DM_TAG_EFI);
> > >>>
> > >>> Under the hood I think for now we could have a simple list of tags for
> > >>> all of DM:
> > >>>
> > >>> struct dmtag_node {
> > >>>     struct list_head sibling;
> > >>>     struct udevice *dev;
> > >>>     enum dm_tag_t tag;
> > >>>     union {
> > >>>        void *ptr;
> > >>>        ulong val;
> > >>>    };
> > >>> };
> > >>
> > >> Just let me make sure; Do you intend that we have a *single* list of tags
> > >> in the system instead of maintaining a list *per udevice*?
> > >
> > > Yes I would prefer not to have a list per udevice, although the API
> > > could be adjusted to iterate through all tags for a particular
> > > udevice, if that is needed (dev_tag_first...() dev_tag_next...().
> >
> > There will never be more than one UEFI handle for one udevice.
> > We need a single field that points to the the handle if such a handle
> > exists. But there will be devices for which UEFI protocols don't exist
> > and where we need no handle. In this case the value can be NULL.
> >
> > Why should we complicate the picture with a list of tags?
> 
> Let's not talk about complexity while we are discussing UEFI :-)
> 
> There are other cases where we need to add info to a device. We cover
> almost all the cases with the uclass-private, plat and priv data
> attached to each device. But in some cases that is not enough,

While I'm not sure whether it is "not enough", I used to think of using
'priv_auto' (or per_device_auto of UCLASS) to hold a pointer to efi_object,
but we might see a conflicting situation in the future where some driver
may also want to use 'priv_auto' for their own purpose.
That is why I added an extra member to udevice.

# The real benefit might be to keep the size of udevice unchanged?

-Takahiro Akashi

> as with
> EFI. I have hit this before in a few other places but have tried to
> work around it rather than extending driver model and adding to the
> already-large struct udevice. But I think we are at the end of the
> road on that.
> 
> I'd also like to look at how much (for example) uclass-plat data is
> used for devices, in case it would be more efficient to move it to a
> tag model.
> 
> I should also point out you are talking about the implementation
> rather than the API. We can always change the impl later, so long as
> we have a suitable API.
> 
> > >
> > > Looking at some of your other patches I think you might need to
> > > support multiple tags for EFI, if there are different things. But
> > > perhaps a list is necesary.
> > >
> > >>
> > >> -Takahiro Akashi
> > >>
> > >>
> > >>> This can be useful in other situations, for example I think we need to
> > >>> be able to send an event when a device is probed so that other devices
> > >>> (with tags attached) can take action. But in any case, it makes the
> > >>> API separate from the data structure, so aids refactoring later.
> > >>>
> > >>> If we find that this is slow we can change the impl, but I doubt it
> > >>> will matter fornow.
> > >>>
> 
> Regards,
> Simon