> Date: Wed, 14 Jul 2021 12:20:08 +0200 > From: Marcus Glocker <mar...@nazgul.ch> > > On Wed, 14 Jul 2021 11:53:50 +0200 (CEST) > Mark Kettenis <mark.kette...@xs4all.nl> wrote: > > > > Date: Mon, 12 Jul 2021 11:56:28 +1000 > > > From: Jonathan Gray <j...@jsg.id.au> > > > > > > On Sun, Jul 11, 2021 at 05:55:14PM +0200, Marcus Glocker wrote: > > > > dwctwo(4) on the Raspberry Pi 3 has some issues today. Basically > > > > uhub2 doesn't work which shows in: > > > > > > > > - mue(4) doesn't attach on uhub2 port 1. > > > > - Removable devices don't attach on uhub2 port 2 and 3. > > > > > > on the 3b+ you mean, smsc(4) is fine on the 3b > > > > > > > > > > > On NetBSD this works fine, but looking for specific fixes in the > > > > NetBSD code base shows quickly that we have grown a big gap > > > > between both code bases. This makes syncing back specific fixes, > > > > if you can identify them at all, very difficult, if not > > > > impossible. > > > > > > > > Therefore I took the path of syncing the entire NetBSD code base > > > > with ours. This fixes the above issues on my Raspberry Pi 3 > > > > which is kind of nice, especially since we can use the integrated > > > > mue(4) Ethernet controller by default now. > > > > > > > > I decided to keep the list_* queue macros from Linux, and use the > > > > Linux headers which are already included in our code base for > > > > drm(4). This just makes future syncing easier. > > > > > > You should not do this. They exist because there are millions of > > > lines of rapidly changing code in drm and only have the barest of > > > what is required for drm. dwc2 is quite small in comparison and > > > I'm sure there are still problems in the drm replacement linux > > > functions. > > > > When I looked at the dwc2 code in the past I learned that replacing > > the Linux list APIs with the interfaces from <sys/queue.h> is not > > trivial. IMHO the Linux list APIs are rather dangerous as they make > > doing list validation almost impossible but they allow some operations > > that the BSD interfaces don't implement. So I think switching back to > > the Linux APIs in this codebase makes sense. But the code should > > probably use its own copy of those interfaces to avoid any unwanted > > interaction with future drm updates as Linux really doesn't have > > stable APIs. > > I'm glad you are highlighting this again, since looking once more at the > porting of the list_* macros just gave me an headache. As you > mentioned correctly, there are certain operations of the list_* macros > which our queue macros don't support, and in the past dwctwo(4) > introduced things like "linked" and "in_freelist" variables to > workaround that, which isn't nice IMO, and turn the code base out of > sync again. > > In my initial version of the synced code I used local copies of the > Linux headers located in sys/dev/usb/dwc2. How about importing the > NetBSD ported Linux headers which are required to sys/dev/usb/dwc2 from > sys/external/bsd/common/include/linux? > > http://cvsweb.netbsd.org/bsdweb.cgi/src/sys/external/bsd/common/include/linux/?only_with_tag=MAIN > > I can shot an adapted diff for that shortly.
Something like that might make sense since that is the implementation that the NetBSD code was tested against. I would prefer if you just imported the bits that you actually need, so maybe cut down the list.h file to size and avoid the other header files if you can.
