I had this talk with Marcel few weeks back about compiling connman on Android/Brillo. This would surely give connman the penetration into Android world and future IoT devices which are based of Brillo. I actually got it compiled and working. There are two challenges: a) Some of the libraries that are required by connman are missing. For my work I got those libraries compiled first. b) Changes in connman itself: Open source connman does not compile as such for Android/Brillo. There are some changes needed in connman to get this going. I would like to see if we can upstream these changes. The intent of this email is to discuss those changes. Following are the changes needed: 1) Backtrace system call: In log.c from function print_backtrace we call backtrace. It is a glibc API and is no present in bionic. There is an equivalent function in bionic for doing this.

2) hard check for ns_initparse in libresolv.so: There is no libresolv.so for android/Brillo. All these ns_* symbols are part of libc.so. So hard check has to go. We may need to check in configure.ac whether we are compiling against glibc or bionic early on and do stuff differently.

3) NTP code calls adjtime to adjust clock. This function is missing in bionic. We may need to find a similar function in bionic for doing this.

4) struct in6_pktinfo is defined in connman code (file gdhcp/common.h) under #ifndef __USE_GNU. Now this structure is defined in bionic header file libc/kernel/uapi/ipv6.h. Not sure what is the intent of #ifndef __USE_GNU? Should this go away? Similarly struct in6_ifreq is defined in inet.c file and it is also defined in the same bionic header file. I had to comment out this definition to compile connman.

5) Some of the structures which are needed for connman compilation are defined under #ifdef __USE_BSD in bionic header file. For Bionic __USE_BSD is defined only if _GNU_SOURCE is defined. So for those files in connman we have to define _GNU_SOURCE as the first line in that file. Example is ipv4ll.c. I hope this should be OK.

6) GResolv structure in gresolv.c: struct __res_state is a member for GResolv structure. This structure is defined in resolv.h file in glibc. struct __res_state in bionic requires inclusion of two other header files resolv_private.h and res_private. Also this structure is defined very differently between bionic and glibc. This caused changes in gresolv.c, the way it accesses various members of this structure.

7) Struct sockaddr_in6 defined differently b/w bionic and glibc.

8) ETH_ALEN is not defined in bionic header file

9) Need to include resolv_params.h in resolver.c. This is because "MAXDNSRCH" etc is defined in resolv_params.h in bionic.

10) Needed to remove "-lresolv -lrt" from Makefile.am.

Some of these changes are easily doable if we can have #ifdef ANDROID in code. Through configure.ac we can detect if we are compiling against bionic or glibc and appropriately turn on this flag. Not sure what your thought are on this.

Hi, On Fri, 2016-07-15 at 18:21 -0700, Naveen Singh wrote: > I had this talk with Marcel few weeks back about compiling connman on > Android/Brillo. This would surely give connman the penetration into > Android world and future IoT devices which are based of Brillo. > > I actually got it compiled and working. There are two challenges: > > a) Some of the libraries that are required by connman are missing. > For > my work I got those libraries compiled first. > > b) Changes in connman itself: Open source connman does not compile as > such for Android/Brillo. There are some changes needed in connman to > get this going. I would like to see if we can upstream these changes. > > The intent of this email is to discuss those changes. Following are > the changes needed: > > 1) Backtrace system call: In log.c from function print_backtrace we > call backtrace. It is a glibc API and is no present in bionic. There > is an equivalent function in bionic for doing this. Backtrace printing was moved to its own file, src/backtrace.c, in ConnMan 1.32, so this is just an automake check away. > 2) hard check for ns_initparse in libresolv.so: There is no > libresolv.so for android/Brillo. All these ns_* symbols are part of > libc.so. So hard check has to go. We may need to check in > configure.ac > whether we are compiling against glibc or bionic early on and do > stuff > differently. Yes, some automake magic needed. > 3) NTP code calls adjtime to adjust clock. This function is missing > in > bionic. We may need to find a similar function in bionic for doing > this. The adjtimex patch on the ml will probably solve this. I'll take a look at that patch nextish. > 4) struct in6_pktinfo is defined in connman code (file > gdhcp/common.h) > under #ifndef __USE_GNU. Now this structure is defined in bionic > header file libc/kernel/uapi/ipv6.h. Not sure what is the intent of > #ifndef __USE_GNU? Should this go away? Similarly struct in6_ifreq is > defined in inet.c file and it is also defined in the same bionic > header file. I had to comment out this definition to compile connman. See commit message for e8784053a0a680ec8943ff20559c6561cb5b9a96. For struct in6_ifreq it looks like very old code that should be fixed. > 5) Some of the structures which are needed for connman compilation > are > defined under #ifdef __USE_BSD in bionic header file. For Bionic > __USE_BSD is defined only if _GNU_SOURCE is defined. So for those > files in connman we have to define _GNU_SOURCE as the first line in > that file. Example is ipv4ll.c. I hope this should be OK. I'd like to keep the files as clean as possible. Looking at autoconf magic, it seems to be possible to use AC_DEFINE to add values to the generated config.h file, which is included first. config.h is used everywhere and should be the first file included. So could one add the needed defines in order to config.h with AC_DEFINE instead? And/or the use of bionic, if needed? > 6) GResolv structure in gresolv.c: struct __res_state is a member for > GResolv structure. This structure is defined in resolv.h file in > glibc. struct __res_state in bionic requires inclusion of two other > header files resolv_private.h and res_private. Also this structure is > defined very differently between bionic and glibc. This caused > changes > in gresolv.c, the way it accesses various members of this structure. Ok, sounds a bit more complicated. You should perhaps post an RFC patch for this so we have an idea of how complicated it becomes. > 7) Struct sockaddr_in6 defined differently b/w bionic and glibc. What is the difference here? Something that does not compile on both "platforms"? > 8) ETH_ALEN is not defined in bionic header file Should be easy once we know bionic is being used. > 9) Need to include resolv_params.h in resolver.c. This is because > "MAXDNSRCH" etc is defined in resolv_params.h in bionic. Automake/conf magic, I hope. > 10) Needed to remove "-lresolv -lrt" from Makefile.am. As above, I hope. > Some of these changes are easily doable if we can have #ifdef ANDROID > in code. Through configure.ac we can detect if we are compiling > against bionic or glibc and appropriately turn on this flag. Not sure > what your thought are on this. I'd go for configure.ac, it makes the code cleaner and uses autoconf/automake for it's intended purpose. Let's start with the easy parts first. Thanks for the heads up!

Patrik

Hi Naveen, > > I had this talk with Marcel few weeks back about compiling connman on > > Android/Brillo. This would surely give connman the penetration into > > Android world and future IoT devices which are based of Brillo. > > > > I actually got it compiled and working. There are two challenges: > > > > a) Some of the libraries that are required by connman are missing. > > For > > my work I got those libraries compiled first. > > > > b) Changes in connman itself: Open source connman does not compile as > > such for Android/Brillo. There are some changes needed in connman to > > get this going. I would like to see if we can upstream these changes. > > > > The intent of this email is to discuss those changes. Following are > > the changes needed: > > > > 1) Backtrace system call: In log.c from function print_backtrace we > > call backtrace. It is a glibc API and is no present in bionic. There > > is an equivalent function in bionic for doing this. > > Backtrace printing was moved to its own file, src/backtrace.c, in > ConnMan 1.32, so this is just an automake check away. > > > 2) hard check for ns_initparse in libresolv.so: There is no > > libresolv.so for android/Brillo. All these ns_* symbols are part of > > libc.so. So hard check has to go. We may need to check in > > configure.ac > > whether we are compiling against glibc or bionic early on and do > > stuff > > differently. > > Yes, some automake magic needed. > > > 3) NTP code calls adjtime to adjust clock. This function is missing > > in > > bionic. We may need to find a similar function in bionic for doing > > this. > > The adjtimex patch on the ml will probably solve this. I'll take a look > at that patch nextish. > > > 4) struct in6_pktinfo is defined in connman code (file > > gdhcp/common.h) > > under #ifndef __USE_GNU. Now this structure is defined in bionic > > header file libc/kernel/uapi/ipv6.h. Not sure what is the intent of > > #ifndef __USE_GNU? Should this go away? Similarly struct in6_ifreq is > > defined in inet.c file and it is also defined in the same bionic > > header file. I had to comment out this definition to compile connman. > > See commit message for e8784053a0a680ec8943ff20559c6561cb5b9a96. > For struct in6_ifreq it looks like very old code that should be fixed. > > > 5) Some of the structures which are needed for connman compilation > > are > > defined under #ifdef __USE_BSD in bionic header file. For Bionic > > __USE_BSD is defined only if _GNU_SOURCE is defined. So for those > > files in connman we have to define _GNU_SOURCE as the first line in > > that file. Example is ipv4ll.c. I hope this should be OK. > > I'd like to keep the files as clean as possible. Looking at autoconf > magic, it seems to be possible to use AC_DEFINE to add values to the > generated config.h file, which is included first. config.h is used > everywhere and should be the first file included. So could one add the > needed defines in order to config.h with AC_DEFINE instead? And/or the > use of bionic, if needed? > > > 6) GResolv structure in gresolv.c: struct __res_state is a member for > > GResolv structure. This structure is defined in resolv.h file in > > glibc. struct __res_state in bionic requires inclusion of two other > > header files resolv_private.h and res_private. Also this structure is > > defined very differently between bionic and glibc. This caused > > changes > > in gresolv.c, the way it accesses various members of this structure. > > Ok, sounds a bit more complicated. You should perhaps post an RFC patch > for this so we have an idea of how complicated it becomes. > > > 7) Struct sockaddr_in6 defined differently b/w bionic and glibc. > > What is the difference here? Something that does not compile on both > "platforms"? > > > 8) ETH_ALEN is not defined in bionic header file > > Should be easy once we know bionic is being used. > > > 9) Need to include resolv_params.h in resolver.c. This is because > > "MAXDNSRCH" etc is defined in resolv_params.h in bionic. > > Automake/conf magic, I hope. > > > 10) Needed to remove "-lresolv -lrt" from Makefile.am. > > As above, I hope. > > > Some of these changes are easily doable if we can have #ifdef ANDROID > > in code. Through configure.ac we can detect if we are compiling > > against bionic or glibc and appropriately turn on this flag. Not sure > > what your thought are on this. > > I'd go for configure.ac, it makes the code cleaner and uses > autoconf/automake for it's intended purpose. Let's start with the easy > parts first. > > Thanks for the heads up! > > > Patrik > Something like this: (it may not be syntactically correct) > with_android=no > AC_MSG_CHECKING([checking for android]) > case "${host}" in > > *android*) > with_android=yes > ;; > > *) > AM_CONDITIONAL([BIONIC], [test "${with_android}" = "yes"]) this should be the last resort in handling it. BlueZ for Android can compile either with glibc/musl or with Bionic. And I think we should get to the level where we do not have huge difference. That said, I recognize that Bionic is just plain ancient and missing anything modern. Which is the biggest pain. Especially with networking capabilities it seems it never caught up to anything recent. I would actually try to detect every missing piece individually and not just use a big hammer like HAVE_BIONIC. With that in mind, lets just get patches to the mailing list and together figure out a way to make this least intrusive to the code base. It is not the first time we dealt with Android specific, but this might be the most painful one since ConnMan uses a lot of low-level stuff. Regards Marcel