Warner Losh
2017-10-09 05:45:37 UTC
I'd like to deorbit /usr/lib/stand.a and /usr/include/stand.h. These are
really parts of the boot loader with an unstable API and shouldn't be
installed into the system. It's really a private library to the boot loader.
Further, I'd like to move the whole shooting match into sys/boot/libsa (the
name change is on purpose to break anything that used libstand).
With these changes, one can cd sys/boot and do a make clean; make all and
have it not depend (too much) on the host computer. It will also let us
clean up a lot of the cut and paste and only half working implementations
of this idea we have now in a clean way.
We can also move the files in sys/boot/common that aren't actually part of
/boot/loader into libsa since there's would be no good reason to keep them
separate (though ultra-tiny loaders can cherry pick individual files, or
compile things in a custom way if that's really needed).
I've been cleaning up the sys/boot Makefile technical debt and this will go
a long way towards disentangling things (though there's a lot to do besides
these fixes).
History is against FreeBSD's practice, btw. We're the only BSD that
cross-threads like this (dragonfly may also, I haven't checked, but they
got it from us if so). NetBSD/OpenBSD have sys/lib/libsa for this stuff.
4.4bsd had each arch build its own libsa.a from files living in
sys/ARCH/stand. 4.2BSD did install things as /usr/lib/libsa.a, but it was
from inside of sys/stand and friends. v7 had a similar structure to 4.2BSD,
but with files living in different directories (they installed libsa.a into
/usr/lib, but again from sys directories). Since we've ditched the 4.4BSD
structure that NetBSD/OpenBSD evolved, I don't think it makes sense to go
the sys/lib route (since this is only used in the boot loader) Since
sys/boot programs know its internals so much that claims of independence
would be hard to sustain. Having it in sys/boot would also help the POLA
this change might cause by fixing a POLA breakage now (it fixes when
libstand API changes break the naive build, as well as ensuring hacks to
libstand will get built when a new boot loader is built). So we invented
the current situation, we can change it since nothing outside the boot
loader uses it (and if there are out-of-tree examples that use it, they
could be converted to the new structure).
Any objections to my moving forward with this (assume I'll help people fix
any in-flight work this disrupts)? I know the plans are a little vague, but
I was hoping to get rough consensus while I came up with woking code and go
all IETF old-school on the problem.
Warner
really parts of the boot loader with an unstable API and shouldn't be
installed into the system. It's really a private library to the boot loader.
Further, I'd like to move the whole shooting match into sys/boot/libsa (the
name change is on purpose to break anything that used libstand).
With these changes, one can cd sys/boot and do a make clean; make all and
have it not depend (too much) on the host computer. It will also let us
clean up a lot of the cut and paste and only half working implementations
of this idea we have now in a clean way.
We can also move the files in sys/boot/common that aren't actually part of
/boot/loader into libsa since there's would be no good reason to keep them
separate (though ultra-tiny loaders can cherry pick individual files, or
compile things in a custom way if that's really needed).
I've been cleaning up the sys/boot Makefile technical debt and this will go
a long way towards disentangling things (though there's a lot to do besides
these fixes).
History is against FreeBSD's practice, btw. We're the only BSD that
cross-threads like this (dragonfly may also, I haven't checked, but they
got it from us if so). NetBSD/OpenBSD have sys/lib/libsa for this stuff.
4.4bsd had each arch build its own libsa.a from files living in
sys/ARCH/stand. 4.2BSD did install things as /usr/lib/libsa.a, but it was
from inside of sys/stand and friends. v7 had a similar structure to 4.2BSD,
but with files living in different directories (they installed libsa.a into
/usr/lib, but again from sys directories). Since we've ditched the 4.4BSD
structure that NetBSD/OpenBSD evolved, I don't think it makes sense to go
the sys/lib route (since this is only used in the boot loader) Since
sys/boot programs know its internals so much that claims of independence
would be hard to sustain. Having it in sys/boot would also help the POLA
this change might cause by fixing a POLA breakage now (it fixes when
libstand API changes break the naive build, as well as ensuring hacks to
libstand will get built when a new boot loader is built). So we invented
the current situation, we can change it since nothing outside the boot
loader uses it (and if there are out-of-tree examples that use it, they
could be converted to the new structure).
Any objections to my moving forward with this (assume I'll help people fix
any in-flight work this disrupts)? I know the plans are a little vague, but
I was hoping to get rough consensus while I came up with woking code and go
all IETF old-school on the problem.
Warner