This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: signal-handling tweaks for mips/mips64

Alexandre Oliva <aoliva at redhat dot com> writes:

> It was reported to me that ucontext is utterly broken, even in o32
> with a stable 32-bit mips kernel.  Indeed, it doesn't match the
> ucontext structure defined by the kernel at all.  This means that
> programs taking real-time signals in o32 won't be able to extract
> correct information from the mcontext_t, since the kernel puts data in
> there that's in an entirely different format.
> I've looked for any ways in which the current data structures could
> possibly be useful, and didn't find any.  gdb and rda thought they
> were using the register arrays, but it turned out they were using the
> arrays in procps instead.  makecontext(), [sg]etcontext() et al aren't
> implemented on mips, so any uses thereof will just return ENOSYS,
> without messing with the given data structure.  So, I believe it is
> not too late for us to fix it such that it matches the kernel data
> structures.
> While at that, I fixed a number of incompatibilities introduced by
> either differences between kernel headers that we used to include,
> whose contents are different depending on whether asm points to
> asm-mips or asm-mips64.
> With this patch, after some pending kernel patches are checked in, one
> will be able to obtain the correct information from signal handlers in
> all mips ABIs.  With n64, this is already true.  With o32, it works
> with the 32-bit mips kernel, but the mips64 kernel needs a patch to
> implement the proper sigcontext ABI.  n32 still a patch to be
> developed for it to be possible for ucontext to be POSIX-compliant.
> Currently, the kernel uses the same ucontext for n32 and n64, but this
> doesn't work in n32 because uc_link must be a pointer and stack_t must
> contain a pointer and a size_t, whose sizes differ between n32 and
> n64.  I believe Ralf is working on a patch for the kernel to generate
> n32-compliant ucontext when invoking signal handlers in n32 processes.
> The only uncertainty is whether uc_flags will be a 32- or 64-bit value
> in n32; I left it as the latter, just because I already had that in
> place; if it changes, a (simplifying) follow-up patch will be posted.
> However, I wanted to circulate the idea of fixing ucontext_t for o32
> as soon as possible, so I didn't wait for a decision on the exact n32
> ABI.
> Ok to install?

 Andreas Jaeger
  SuSE Labs aj at suse dot de
   private aj at arthur dot inka dot de

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]