| Message ID | mvmtwemqkbi.fsf@hawking.suse.de |
|---|---|
| State | New |
| Headers | show |
On 08/15/2016 02:37 PM, Andreas Schwab wrote: > __startcontext marks the bottom of the call stack of the contexts created > by makecontext. > > [BZ #20435] > * sysdeps/unix/sysv/linux/arm/setcontext.S (__startcontext): Mark > as .cantunwind. > --- > sysdeps/unix/sysv/linux/arm/setcontext.S | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/sysdeps/unix/sysv/linux/arm/setcontext.S b/sysdeps/unix/sysv/linux/arm/setcontext.S > index 603e508..d1f168f 100644 > --- a/sysdeps/unix/sysv/linux/arm/setcontext.S > +++ b/sysdeps/unix/sysv/linux/arm/setcontext.S > @@ -86,12 +86,19 @@ weak_alias(__setcontext, setcontext) > > /* Called when a makecontext() context returns. Start the > context in R4 or fall through to exit(). */ > + /* Unwind descriptors are looked up based on PC - 2, so we have to > + make sure to mark the instruction preceding the __startcontext > + label as .cantunwind. */ > + .fnstart > + .cantunwind > + nop Why isn't the PC - 2 kludge needed for the other .cantunwind cases? Would it be possible to add a C test case to the glibc testsuite? Thanks, Florian
On Aug 15 2016, Florian Weimer <fweimer@redhat.com> wrote: > Why isn't the PC - 2 kludge needed for the other .cantunwind cases? Because the other uses have regular function calls, so they return back to the middle of the function that is marked cantunwind. __startcontext doesn't contain the function call, but is the address the context returns to. > Would it be possible to add a C test case to the glibc testsuite? This is difficult, because a simple call stack won't trigger the bug. Andreas.
On 08/15/2016 03:02 PM, Andreas Schwab wrote: > On Aug 15 2016, Florian Weimer <fweimer@redhat.com> wrote: > >> Why isn't the PC - 2 kludge needed for the other .cantunwind cases? > > Because the other uses have regular function calls, so they return back > to the middle of the function that is marked cantunwind. __startcontext > doesn't contain the function call, but is the address the context > returns to. > >> Would it be possible to add a C test case to the glibc testsuite? > > This is difficult, because a simple call stack won't trigger the bug. Thanks for the explanation. Your patch looks okay to me. Florian
diff --git a/sysdeps/unix/sysv/linux/arm/setcontext.S b/sysdeps/unix/sysv/linux/arm/setcontext.S index 603e508..d1f168f 100644 --- a/sysdeps/unix/sysv/linux/arm/setcontext.S +++ b/sysdeps/unix/sysv/linux/arm/setcontext.S @@ -86,12 +86,19 @@ weak_alias(__setcontext, setcontext) /* Called when a makecontext() context returns. Start the context in R4 or fall through to exit(). */ + /* Unwind descriptors are looked up based on PC - 2, so we have to + make sure to mark the instruction preceding the __startcontext + label as .cantunwind. */ + .fnstart + .cantunwind + nop ENTRY(__startcontext) movs r0, r4 bne PLTJMP(__setcontext) @ New context was 0 - exit b PLTJMP(HIDDEN_JUMPTARGET(exit)) + .fnend END(__startcontext) #ifdef PIC