From mboxrd@z Thu Jan 1 00:00:00 1970 Message-ID: <54D8F09A.7030802@control.lth.se> Date: Mon, 09 Feb 2015 18:38:34 +0100 From: Anders Blomdell MIME-Version: 1.0 References: <54D4D409.2020708@control.lth.se> <54D4D9B2.8090805@xenomai.org> <54D4DF0C.2080007@control.lth.se> <54D4E285.9000906@xenomai.org> <54D4E44B.4060403@control.lth.se> <20150206161215.GC27277@hermes.click-hack.org> <20150206161626.GD27277@hermes.click-hack.org> <54D8D7D5.2060000@control.lth.se> <20150209155716.GE3200@hermes.click-hack.org> <54D8DC0B.3070107@control.lth.se> <20150209162409.GF3200@hermes.click-hack.org> <54D8E83F.6020601@control.lth.se> <54D8EE0A.8040600@xenomai.org> In-Reply-To: <54D8EE0A.8040600@xenomai.org> Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 8bit Subject: Re: [Xenomai] Mixing linux and alchemy (cobalt) calls List-Id: Discussions about the Xenomai project List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Philippe Gerum , Gilles Chanteperdrix Cc: "Xenomai@xenomai.org" On 2015-02-09 18:27, Philippe Gerum wrote: > On 02/09/2015 06:02 PM, Anders Blomdell wrote: >> On 2015-02-09 17:24, Gilles Chanteperdrix wrote: >>> On Mon, Feb 09, 2015 at 05:10:51PM +0100, Anders Blomdell wrote: >>>> On 2015-02-09 16:57, Gilles Chanteperdrix wrote: >>>>> On Mon, Feb 09, 2015 at 04:52:53PM +0100, Anders Blomdell wrote: >>>>>> On 2015-02-06 17:16, Gilles Chanteperdrix wrote: >>>>>>> On Fri, Feb 06, 2015 at 05:12:15PM +0100, Gilles Chanteperdrix wrote: >>>>>>>> On Fri, Feb 06, 2015 at 04:56:59PM +0100, Anders Blomdell wrote: >>>>>>>>> On 2015-02-06 16:49, Philippe Gerum wrote: >>>>>>>>>> On 02/06/2015 04:34 PM, Anders Blomdell wrote: >>>>>>>>>>> On 2015-02-06 16:11, Philippe Gerum wrote: >>>>>>>>>>>> On 02/06/2015 03:47 PM, Anders Blomdell wrote: >>>>>>>>>>>>> I have an application that need both realtime and linux sockets, am I correct in assuming that >>>>>>>>>>>>> withe the alchemy skin I could access them like >>>>>>>>>>>>> >>>>>>>>>>>>> socket(... // Linux version >>>>>>>>>>>>> __real_socket(... // Linux version >>>>>>>>>>>>> __cobalt_socket(... // Alchemy/cobalt version >>>>>>>>>>>>> >>>>>>>>>>>>> while under the cobalt skin, it would be: >>>>>>>>>>>>> >>>>>>>>>>>>> socket(... // Alchemy/cobalt version >>>>>>>>>>>>> __real_socket(... // Linux version >>>>>>>>>>>>> __cobalt_socket(... // Alchemy/cobalt version >>>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> This depends on the LDFLAGS retrieved from xeno-config: >>>>>>>>>>>> >>>>>>>>>>>> 1. with --posix mentioned in the xeno-config --ldflags request >>>>>>>>>>>> >>>>>>>>>>>> socket(...), __cobalt_socket(...) or __RT(socket(...)) => Cobalt >>>>>>>>>>>> implementation >>>>>>>>>>>> __real_socket(...) or __STD(socket(...)) => glibc service >>>>>>>>>>>> >>>>>>>>>>>> 2. without --posix mentioned in the xeno-config --ldflags request >>>>>>>>>>>> >>>>>>>>>>>> __cobalt_socket(...) or __RT(socket(...)) => Cobalt implementation >>>>>>>>>>>> socket(...) or __STD(socket(...)) => glibc service >>>>>>>>>>>> >>>>>>>>>>>> e.g. >>>>>>>>>>>> >>>>>>>>>>>> - the application only wants to access the POSIX services implemented by >>>>>>>>>>>> Cobalt using the regular POSIX names: LDFLAGS should contain the output of: >>>>>>>>>>>> $ xeno-config --posix --ldflags, or --cobalt --ldflags. >>>>>>>>>>>> >>>>>>>>>>>> - the application wants to access the POSIX services implemented by >>>>>>>>>>>> Cobalt using the regular POSIX names, and the alchemy API: LDFLAGS >>>>>>>>>>>> should contain the output of: >>>>>>>>>>>> $ xeno-config --posix --alchemy --ldflags, or --cobalt --alchemy --ldflags. >>>>>>>>>>>> >>>>>>>>>>>> - the application wants to access the POSIX services implemented by >>>>>>>>>>>> Cobalt solely via the explicit POSIX wrappers, and the alchemy API: >>>>>>>>>>>> LDFLAGS should contain the output of (i.e. omitting --posix): >>>>>>>>>>>> $ xeno-config --alchemy --ldflags, or --alchemy --ldflags. >>>>>>>>>>>> >>>>>>>>>>>> NOTE: using __RT() is preferred to calling __cobalt(), in case an API >>>>>>>>>>>> stacked over the Cobalt POSIX API redefines its own implementation of >>>>>>>>>>>> POSIX services over the dual kernel. __RT() would call the stacked >>>>>>>>>>>> implementation, __cobalt() would force a call to the Cobalt >>>>>>>>>>>> implementation of the service. >>>>>>>>>>>> >>>>>>>>>>> Thanks for the clarification, will sprinkle the code with __STD(...) >>>>>>>>>>> and __RT(...), from here on :-). >>>>>>>>>>> >>>>>>>>>> >>>>>>>>>> That's only required if you want your code to unambiguously route to the >>>>>>>>>> proper service in case the default symbol wrapping does not fit, or is >>>>>>>>>> not present. This is typically what libcopperplate does, so that >>>>>>>>>> non-POSIX apps can link against it, without being required to wrap the >>>>>>>>>> POSIX symbols in the final executable. >>>>>>>>> >>>>>>>>> More to avoid me shooting myself in the foot when trying to juggle sockets >>>>>>>>> from two different domains (also makes the code less dependent on the linker >>>>>>>>> flags given). Already got bitten by 'modprobe rtpacket' not loading properly >>>>>>>>> and the __wrap_socket picking up the posix version. And of course making the >>>>>>>>> code clearly document what belongs where. >>>>>>>> >>>>>>>> Not loading rtpacket should cause __wrap_socket to use the posix >>>>>>>> version only if you are trying to create a socket type that the >>>>>>>> rtpacket module implements. Otherwise, this is a bug. >>>>>>> >>>>>>> And unless code has changed between 2.x and 3.x in this area, using >>>>>>> __RT() will result in exactly the same behaviour. >>>>>> >>>>>> You might be right, but AFAICT, on 2.6.2.1 'int __rt_dev_socket(...)' >>>>>> (ksrc/skins/rtdm/core.c) calls 'struct rtdm_device *get_protocol_device(...)' >>>>>> (ksrc/skins/rtdm/device.c), while in xenomai3/next 'COBALT_IMPL(int, socket,...)' >>>>>> does a 'XENOMAI_SYSCALL3(sc_cobalt_socket, ...)' and then does a failover to >>>>>> '_STD(socket, ...)' in case of -ENOSYS (which is what I believe an unloaded rt_packet.ko >>>>>> gives as a result). >>>>>> >>>>>> Am I missing something obvious? >>>>> >>>>> You are comparing user-space code with kernel space code. In xenomai >>>>> 2.6, the user-space code you should be looking at is >>>>> src/skins/posix/rtdm.c >>>> Even from native mode (which is what I use in 2.6)? >>> >>> In that case, you should look at src/skins/rtdm/core.c >>> The services rt_dev_socket belongs to the rtdm skin not to native >>> skin. >> I stand humbly corrected. >> >>> >>>> >>>>> And in that code __wrap_socket falls back to __real_socket if the >>>>> kernel-space code returns one of the known errors. >>>>> >>>>> Now the real question is: what arguments do you pass to the cobalt >>>>> rt socket call? If these arguments are the one corresponding to the >>>>> socket protocol implemented by the rtpacket module, the behaviour >>>>> you observe is normal, otherwise, this is a bug. >>>> I don't say it's a bug, but to me it seems like alchemy differs from >>>> the old native skin in this case. >>> >>> Well, not really, the native skin never had a socket call anyway, >>> and alchemy does not either. It is just that the rtdm skin >>> disappeared, so now you use the cobalt service instead of the rtdm >>> skin. >> # /usr/xenomai/bin/xeno-config --version >> 3.0-rc2 >> # usr/xenomai/bin/xeno-config --help >> Usage xeno-config OPTIONS >> Options : >> --help >> --v,--verbose >> --version >> --cc >> --ccld >> --arch >> --prefix >> --[skin=]posix/cobalt|vxworks|psos|alchemy|rtdm|smokey >> >> --rtdm is there, bug? > > The ambiguous rt_dev_*() API has been obsoleted, since any application > can now use the POSIX call (all apps have libcobalt underneath). Compat > wrappers should be provided by libtrank though, I'll push this before 3.0. > > We will then have: > > - rtdm_open/read/write...() => inter-driver calls, kernel-space only > - open/read/write...() => RTDM service calls for applications > - rt_dev_open/read/write...() => compat wrappers from libtrank to > open/read/write...(), for 2.x apps that need them. Ok, so for libraries where I don't know which skin will be used for linking, the best way would be to use __RT(...) for the calls, thereby forcing the symbol to be whatever the skin --cflags stipulates (if I don't use __RT, the symbol might unwrapped at link time), and then do an ioctl(RTIOC_DEVICE_INFO) to check that I really got a realtime filedescriptor? /Anders -- Anders Blomdell Email: anders.blomdell@control.lth.se Department of Automatic Control Lund University Phone: +46 46 222 4625 P.O. Box 118 Fax: +46 46 138118 SE-221 00 Lund, Sweden