On Sun, May 06, 2018 at 04:04:02PM +0100, Peter Maydell wrote:
> On 6 May 2018 at 14:39, David Gibson <david@gibson.dropbear.id.au> wrote:
> > Well, I'm biased of course, but I think we'd be better off just
> > ditching the wrapper.  In its present form it is so limited as to not
> > really add any value.  If it was rewritten to do something useful
> > (e.g. handling reallocations), I think it would be even better if
> > done as an extension to libfdt itself so it can benefit everyone, not
> > just qemu.
> 
> Well, some of it is working around infelicities in libfdt's
> API (like all the getprop/setprop functions taking an offset
> value rather than a node name), but yes, it would be better
> to fix the libfdt API if possible.
> 
> > Although, that said, I'll re-iterate that I think qemu's fdt
> > manipulation is now sufficiently complex that it would be better off
> > using a "live" (dynamically allocated & pointer based) tree
> > representation that we just flatten immediately before loading it into
> > the guest.
> 
> This sounds to me like something that should be handled
> by libfdt.

No, it's really not.  libfdt is specifically for reading and writing
flattened trees in place, in flattened form.  That's the whole basis
of the design and it's directly responsible for some of the
infelicities you mention.

I was never intended as a general runtime t manipulation library.  A
libdt that does that (using an allocator and internal pointer
representation) would be a nice thing to have, but libfdt isn't it.
Obviously being able to use libfdt to flatten and unflatten trees
would be a very good feature for such a library.

> internal representation of the data structure is, I just
> want to be able to (a) hand it an fdt read in from a file
> (b) call various functions to modify the data structure
> and then (c) write the resulting thing out to an fdt in
> guest memory. Whether libfdt prefers to do that by
> modifying the flat representation or by converting into
> a dynamically allocated unflattened tree and back again
> is something I'd rather leave to it as an implementation
> detail.

Sort of.  But there design constraints in libfdt which means that's
not really feasible.  Such as:
  * not requiring an allocator
  * not requiring a "read / unflatten" pass before reading values from
    an fdt
  * not requiring allocation or creation of a "context" state for
    manipulating a tree

Those aren't aren't really relevant to qemu and perhaps aren't part of
what you'd usually think of as the API.  But they do matter to real
libfdt use cases.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson