Re: Time to deprecate crash dump support in stats tools (or KVM-grovelers in general)?

[David Holland <dholland-tech%netbsd.org@localhost>]

On Mon, Apr 14, 2008 at 03:15:08PM +1000, Simon Burge wrote:
 > > For debugging, I think what's wanted is one tool that reads the
 > > complete sysctl/kernfs/*ctl/whatnot tree from a crash dump, and set up
 > > tools like netstat to be able to read from that. Or maybe even create
 > > a wedge in the libc sysctl interface for it. This would allow almost
 > > anything that can be done now, only more reliably.
 > 
 > The problem with this is that a number of sysctl's aren't just "read"
 > from the kernel - the kernel does some amount of work to build the reply
 > first then copies the results out to userland.  Common examples are
 > things that follow linked lists in the kernel, but then present a nice
 > simple array (often nice 64-bit alignment friendly structures instead of
 > "raw" structures used in the original data) to userland.

Sure, but kmem grovelers already have to do this. Maybe they don't do
it in quite as general a fashion... but there's no real reason they
couldn't. And if they did, it would probably be feasible to generate
or share the code with the kernel instead of maintaining multiple
copies by hand.

Not that this is entirely trivial.

-- 
David A. Holland
dholland%netbsd.org@localhost