Re: [libunwind] backtrace on EM64T

[Keith Owens]

On Mon, 21 Nov 2005 15:55:35 -0800, 
Arun Sharma <address@hidden> wrote:
>David Mosberger-Tang wrote:
>
>>>I can't seem to find a noticable improvement.
>> 
>> 
>> That's because caching has not been implemented yet for DWARF2 (it is,
>> for the ia64 unwinder and makes a huge difference).  The code is
>> already set up to handle caching, but somebody would have to fill in
>> the actual code...
>
>I see that ia64 implements a script cache and there is no such thing on 
>x86-64. Would it be correct to say that a dwarf2 based system should 
>implement a cache of ip -> unw_proc_info_t mappings?

One problem with the ia64 unwind cache is that it uses kmalloc().  Code
expects to be able to call the unwinder when the system is running out
of memory or even from an NMI/MCA/INIT handler, at which point the slab
locks and chains are unreliable.  My KDB patches introduce a last ditch
allocator for the ia64 unwinder, so we can get backtraces even when the
entire system is dying.