On 07/24/2015 09:24 AM, Philippe Gerum wrote: > On 07/23/2015 09:45 PM, Philippe Gerum wrote: >> On 07/23/2015 03:37 PM, Jan Kiszka wrote: >>> On 2015-07-23 11:22, Philippe Gerum wrote: >>>> On 07/23/2015 11:08 AM, Jan Kiszka wrote: >>>>> On 2015-07-23 10:40, Philippe Gerum wrote: >>>>>> On 07/21/2015 04:54 PM, Jan Kiszka wrote: >>>>>>> On 2015-07-21 14:51, Philippe Gerum wrote: >>>>>>>> On 07/21/2015 02:27 PM, Jan Kiszka wrote: >>>>>>>>> Hi Philippe, >>>>>>>>> >>>>>>>>> just a heads up, I'll try to address this later: >>>>>>>>> >>>>>>>>> I received a report that include/rtdm/uapi/rtdm.h contains some data >>>>>>>>> structures that are used between the kernel and the userspace library >>>>>>>>> (so not directly by applications) and are not compatible with >>>>>>>>> 32-on-64-bit (compat) scenarios. We should probably promote anything >>>>>>>>> that is longer on 64-bit to that size, unconditionally. As this >>>>>>>>> affects >>>>>>>>> the ABI, it should be fixed before 3.0 release, ideally. >>>>>>>>> >>>>>>>> >>>>>>>> That would require to fix up the client drivers (setsockopt, getsockopt >>>>>>>> typically). Ok, I'll have a look when I'm done with testing the >>>>>>>> blackfin >>>>>>>> port. >>>>>>> >>>>>>> Indeed. Just realized that there are already compat structs for the >>>>>>> socket stuff. However, and that was what our user stumbled over, there >>>>>>> is none for _rtdm_mmap_request. But that should be really internal, >>>>>>> right? >>>>>>> >>>>>> >>>>>> mmap() has the compat thunk for 32<->64 bit conversion already, so this >>>>>> is ok. >>>>> >>>>> Ok... Rechecking the bug report here, the issue is rather the case that >>>>> off_t becomes 64 bit when defining -D_FILE_OFFSET_BITS=64 for your >>>>> userland. We are apparently required to account for that case as well. >>>>> >>>> >>>> Ok, so let's move the offset member to loff_t and compat_loff_t in the >>>> mmap request descriptors, that should do the trick. >>>> >>> >>> I only forwarded this suggestion internally, unfortunately cannot tell >>> yet how it was implemented, just received the early feedback that it is >>> not yet working. Maybe we are overseeing something, maybe something went >>> wrong with the implementation. Will be checked tomorrow. >>> >> >> This is likely an issue with the native alignment of those types, the >> kernel may offset long members in the compat struct based on different >> constraints than -m32 does. >> > > And we probably also need the mmap64 wrapper we don't have yet. Ok, I'm > working on this. >
Pushed to -next, and tested over 32 and 3264 ABI models over x86. This breaks the ABI, so both kernel and userland have to be rebuilt. The bad news is that I now have to redo all the testing on all other archs. Oh well, ... -- Philippe. _______________________________________________ Xenomai mailing list [email protected] http://xenomai.org/mailman/listinfo/xenomai
