Description of problem:
The new version of ia32-el does not set "current->personality" to
PER_LINUX32. Our 64-bit kernel modules rely on this to determine
whether the calling process is 32-bit or 64-bit. This was getting
set in earlier RHEL4 drops.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1. build a 32-bit app that calls an ioctl into a 64-bit kernel module.
2. run the 32-bit app on IA64 using the ia32EL.
3. check inside the kernel that current->personality is not set to
current->personality is not set.
current->personality = PER_LINUX32 for 32-bit calling processes.
Is this a bug, or an intended change. If intended, what will be the
new method for determining whether the calling process is 32-bit or
64-bit? Ideally, we'd like this to be the same across all 64-bit
platforms (ie: on x86_64 we also check current->personality).
I doubt ia32el calling personality (PER_LINUX32) will make any difference
on RHEL4 - as the kernel is compiled without CONFIG_IA32_SUPPORT, that
personality is simply not registered, so likely the syscall will fail.
I believe the right solution would be to add an interface for external ioctl/syscall translation modules to ia32el and let say Veritas build/ship
an ia32el module that would do the translation.
How should Veritas test for 32-bit process is a question for the kernel team,
but testing PER_LINUX32 sounds wrong, one can change the personality.
On x86-64, the right way is IMHO test_thread_flag(TIF_IA32).
On IA-64 with ia32el, this is very questionable, as the process is really 64-bit,
*** Bug 139698 has been marked as a duplicate of this bug. ***
*** Bug 139567 has been marked as a duplicate of this bug. ***
I also expected that a "personality(PER_LINUX32)" call would fail on
RHEL4 ... but I tried it and it works just fine. So this does
provide a simple and convenient way for ia32-el to leave a footprint
that an in-kernel module can test to determine how to interpret the
format of a binary structure passed between usermode and kernel.
Ia32-el can't make the conversion ... it has no idea what the
structures look like. A plug-in interface sounds complex/cumbersome
(some sort of extra shared libraries that ia32-el would load if they
are present, and then feed all ioctl's through??)
I was wrong, personality (PER_LINUX32) in ia32el-1.1-11 succeeds
(the difference from CONFIG_IA32_SUPPORT times is just that this call doesn't
change exec domain and therefore doesn't change e.g. altroot).
So I guess checking current->personality == PER_LINUX32 is something that
Veritas can use in their module.