On 07/26/2017 02:52 PM, Michael S. Tsirkin wrote:
On Tue, Jul 18, 2017 at 06:34:37PM +0200, Marc-André Lureau wrote:
...
In the future, we may want to hide vhost-user from QAPI/introspection
with conditional compilation, although the design of this hasn't been
fully fleshed out yet and shouldn't prevent this patch from being
applied.
Signed-off-by: Marc-André Lureau <marcandre.lur...@redhat.com>
After an offline discussion, looks like the idea is to be able to ship a
cut down binary with less features (and e.g. smaller attack surface?).
I think it deserves it own thread, there will be many ideas about it.
I liked this idea a first because it speeds up compilation time (so the
CI build system can run more tests, cover more of less code...) but then
we have this unreadable unmanageable ifdefery very hard to maintain, you
end up not compiling this piece of code, then move things around,
refactor a bit, change some API and zombie code survives around until
someone who has this lib try to compile the obsolete code.
This is also why I now prefer to build via docker images (the
qemu:debian-amd64 try to come with as many dev packages as possible to
increase code coverage).
There is some dream about having one monolithic QEMU binary able to run
all cpus/machines/devices. I think this way do reduce attack surface by
exposing all the surface. This is a bit what is happening during release
candidate testing when /master is closed, people start to use QEMU
[differently than while adding features], more surface is stressed, then
many issues appear and get resolved.
Thomas Huth did some effort adding more poisoned macros and move common
objects to common-obj, we can do more here, we may need to split more
sources to move more common code.
Another experience is I'm trying to model a 32-bit big endian ARM board
then got bugged with 64-bit code this board will never use, or issue
with PCI and AHCI, so I started to comment those objects using
obj-$(CONFIG_X) but then CONFIG_X also depends of CONFIG_Y and so, then
it become hardly scalable, see [1]:
+CONFIG_LAN9118=$(or
$(CONFIG_REALVIEW),$(CONFIG_EXYNOS4),$(CONFIG_FSL_IMX31))
+CONFIG_SMC91C111=$(or
$(CONFIG_PXA2XX),$(CONFIG_MAINSTONE),$(CONFIG_REALVIEW))
+CONFIG_ARM_MPTIMER=$(or
$(CONFIG_A9MPCORE),$(CONFIG_ARM11MPCORE),$(CONFIG_A15MPCORE))
+CONFIG_HIGHBANK=$(and $(CONFIG_A9MPCORE),$(CONFIG_A15MPCORE))
we already have:
crypto-obj-$(if $(CONFIG_NETTLE),n,$(if $(CONFIG_GCRYPT_HMAC),n,y)) +=
hmac-glib.o
crypto/Makefile.objs:23:crypto-obj-$(if $(CONFIG_GCRYPT),n,$(if
$(CONFIG_GNUTLS_RND),n,y)) += random-platform.o
I ended dropping this branch, I was thinking about using kconfig to
manage dependencies... The obj-$(CONFIG_XYZ) is useful when kept tiny,
at most 1 bitwise operation.
I wanted to share my experience and thoughts about it, I hope it's still
understandable because it's very late :S
Regards,
Phil.
> Besides failing build on freebsd, this patch has too much ifdefery for
> my taste. How about we patch just net/net.c?
[1]
https://github.com/philmd/qemu/commit/2c1c515345d626f47d0968821980f9389d4588b2#diff-6006132695caf8d409e781fbf3d4401f
