Thanks Philippe for your quick reply. May be a few additional remarks will clarify some remaining misunderstanding. Compared to Xenomai the basic differences are -- there are no shadow threads -- we use the standard glibc with futex based synchronization/communication working across domain boundary transparently. This transparency is demonstrated by having the process binary running on a ipipe-patched Linux only. After loading the realtime module the same binary runs but the responding threads provide for the realtime guarantees. Now lets go beyond marketing: First the glibc implementation is not completely realtime capable. This concerns two functions - the implementation of the SIGEV_THREAD notification - the implementation of the spinlock() function For both we do have realtime capable implementations. Since these issues hold also for a natively realtime capable Linux it also must be solved for a "natively" realtime capable Linux (PREMPT_RT) and the chances to have one standard implemention are quite good. The only difference may be the spinlock(). Here the protection needed for static priorities can be done differently. For a two kernel solution the protection must work across (interrupt disable).
I guess the transparency also explains why we cannot rely on lazy migration back. If for e.g. a thread which should provide for realtime response does an open() or a mmap() during its setup phase and then a sigwait() for responding, then the sigwait() call has to be executed in the realtime domain from the very beginning. Having the check for migration back to realtime at the system call epilog of Linux is the most convenient way, otherwise we would neet hooks in every system call function which is propagatable. Concerning the futex function. Currently we intercept at the system call exit and call the corresponding rt-function when the number of requested wakeups could not be performed. This provides for an excellent filtering but works for regular mutexes only. If we want to preserve the exact semantics for PI mutexes we have to call the rt-function upfront. For mutexes with priority ceiling migration the migration check at system call exit is sufficient. For priority inheritance we would need to use the scheduler hook. Concerning the mlock-stuff we view it to be not sufficient, since if somebody does a malloc() and sets up preallocated structures, they are not necessarily touched. Concerning the performance issues and your remark that you have still work to do. For us treating system calls what they really are namely synchronous exceptions which should be handled by the causing domain only would be perfect fit and would be faster.(as an option) Hopefully this clears up the issues somewhat. Karl-Heinz -----Ursprüngliche Nachricht----- Von: Philippe Gerum [mailto:[EMAIL PROTECTED] Gesendet: Mittwoch, 7. Juni 2006 15:21 An: Krause, Karl-Heinz Cc: email@example.com Betreff: Re: [Xenomai-core] Ipipe hook at system call exit Hello, Krause, Karl-Heinz wrote: > Hallo Philippe > > > > Jan Kiszka referred me to you discussing our problem with a missing > Ipipe hook at system call exit. > > We at Siemens A&D do have a Linux realtime approach which is based on a > previous ADEOS version. When trying to port an improved version to the > Ipipe version for kernel 188.8.131.52 we ran into the problem of not having > an event hook at system call exit. Let me explain the need for it by > briefly outlining our approach. > > It is a two kernel approach based on the model of a multihreaded process > (means 2.6 kernel) where the threads above a certain static priority > level e.g. 68 are scheduled by the scheduler of the realtime kernel. > The realtime kernel maintains exactly the same systemcall interface as > the Linux kernel. The entire process works uniformely with the glibc. > The glibc isn't aware under which scheduler the current thread is > executing. To make this happen and having both schedulers to work with > the same struct task struct we had to put some restrictions on the > signalling for the realtime domain (restrictions which make sense for > the realtime arena anyway). Because of that transparency this approach > combines somehow the advantages of a separated realtime kernel with the > user convenience of PREEMPT_RT. (the user convenience was the driving > requirement for our approach) > There seems to be quite a lot of commonality with the way Xenomai deals with shadow threads to enable realtime processing in user-space, while providing a seamless integration with Linux. One difference might be the way your system deals with Linux syscalls fired on behalf of a thread controlled by the real-time scheduler; Xenomai migrates the thread to the Linux scheduler transparently, but I did not figure out yet if this was a relevant issue in your system. Anyway, I think that I now roughly understand the general dynamics of it, thanks for the explanations. > > > Now to the question why we need a hook at systemcall exit. > > The hook at systemcall entry branches to the system call handling of the > realtime kernel, which is also entered via a systemcall table. The > handling can be grouped in three classes > > - complete handling in the realtime domain e.g. timer_settime(), > sigwait() > > - only migration of the thread to the Linux scheduler. Basically > all calls needed for setup e.g. open(), mmap(), pthread_create(). The > migration is transparent for the ipipe code, the thread continues > execution in the Linux domain with the call of the Linux system call > table (the priority hasn't changed). > > - handling in the realtime domain and migration to the Linux > domain if the thread priority has dropped unter the boundary (e.g > releasing a mutex with priority ceiling) > > > > In particular for the second case a check needs to be done at sytem call > exit as to whether the thread has to migrate (back) to the realtime > scheduler. But this is also needed when a call issued in the Linux > raises the priority above the threshold. A third reason for the hook is > to touch the corresponding pages after a brk() or mmap() call for > getting residency. > > Note: > > The migration only takes place for threads of a process marked as realtime. > > Currently we allow only for one realtime process. First it is sufficient > for us and second it allows us to maintain the futex queue (each domain > maintains a local queue) of the realtime domain with virtual addresses > (no mm_lock). > Does this mean that you specifically intercept futex ops to process them in real-time mode when fired over the real-time context? Which would in turn allow you to traverse most of the glibc code and get it synchronized with the plain Linux threads? > > > So this hook at system call exit is a necessity for us. Of course we > could do a private patch, but do you see a possibility to have it in the > standard Ipipe-patch? > Basically, I removed the sysexit hook from the I-pipe patch because it added a non-negligible overhead to each syscall. Even the sysenter hook needs some work to reduce its CPU footprint and I've planned to tackle the issue soon. For this reason, the current Xeno implementation only relies on the sysenter (IPIPE_EVENT_SYSCALL) hook to deal with migrations between the Linux and Xenomai schedulers, usually enforcing a lazy migration scheme, i.e. the syscall prologue added by the RT extension switches the caller to the proper domain before running the system call handler, but does not eagerly switch back to the originating domain (well, there are exceptions to this, but that's the usual way things are handled). Reading your description, a few questions came to my mind: - why do you force a switch back to the originating domain? IOW, are eager transitions absolutely required in your design, since your RT thread is underlaid by a regular Linux task anyway, so it could continue its processing and switch back to the RT side only when needed? - would not it be possible to intercept the IPIPE_EVENT_SETSCHED notifications, which are fired by the I-pipe when a Linux task is about to have its priority changed? It's a direct hook from the kernel's sched_setscheduler(), which is given the task_struct pointer of the altered task, right after its priority field has been updated, but still before the Linux runqueue is reordered. - would mlocking the data segment of your application be enough/possible to ensure that brk() and mmapped() segments get committed to physical memory automatically, and as such spare you the need for touching those areas explicitely? AFAIK, mlocked pages are going to be fixed up this way by the mm layer during the mlocking call. - generally speaking, since you control the prologue and epilogue of all system calls (Linux or real-time) which go through your own syscall demux by mean of the IPIPE_EVENT_SYSCALL hook, it should be possible to handle the whole migration issue (be it eager or lazy in this case) from your code, instead of relying on a hook inserted in Linux's syscall return path. Or am I missing something? > > > > > Karl-Heinz Krause > > Siemens A&D > > Nbg.-Moorenbrunn > > > > > > > > > > > ------------------------------------------------------------------------ > > _______________________________________________ > Xenomai-core mailing list > Xenomaifirstname.lastname@example.org > https://mail.gna.org/listinfo/xenomai-core -- Philippe. _______________________________________________ Xenomai-core mailing list Xenomaiemail@example.com https://mail.gna.org/listinfo/xenomai-core