[PATCH 0/2] Global breakpoints, introduction

[Stan Shebs <stanshebs at earthlink dot net>]

This pair of patches plus Linux kernel module adds global breakpoints to 
GDB.  Global breakpoints are installed in multiple processes of a 
target, *including* processes not under GDB control, and processes not 
yet created.  Conceptually, it's similar to breakpoints in some RTOSes 
where tasks share all or part of an address space, and so a breakpoint 
in the shared space will hit in whatever task reaches it first.

In Linux, this is a little more complicated.  We don't want to force 
every process to run under the debugger all the time, and we need to 
catch new processes somehow.  The implementation here uses a kernel 
module that communicates with debuggers via a device /dev/breakpoint.  
In the absence of appropriate internal support in Linux, the module uses 
kprobes to hook into several key spots, and then installs and manages 
breakpoint traps on behalf of the debugger.  When one of these traps is 
hit, the device removes the trap and notifies GDB, which then attaches 
to the process and installs its own breakpoint trap in the usual way.

The user interface adds a "process" keyword to breakpoints, along the 
lines of "thread", so for instance "break mysubr process *" installs a 
breakpoint in all processes using the file in which mysubr is defined 
(which could be either the exec file or a shared library).  "break 
mysubr process future" installs only in processes created after the 
breakpoint is defined.  In addition, combinations and process user/owner 
qualifications are possible, so you can do instant system catatonia with 
something ill-advised like "break malloc process all/root". :-)

The patch here includes both native Linux and remote protocol for Linux 
gdbserver, and I've attached a tgz of the (small) kernel module sources 
to this message.  (The module has only been tested much on the 2.6.35 
kernel.)

This is kind of a freaky addition - it's very cool to see in action, as 
you can catch the first process to call some library function, and it 
will be even more interesting for multicore situations where lots of 
copies of a exec may be running.  At the same time, it pushes the 
envelope on hooking into the system, and in the absence of full 
multi-process infrastructure, the user interface is a little hacky.  So 
I'm open to ideas on how to do all this better.

Stan
stan@codesourcery.com

Attachment: bpmodule.tgz
Description: GNU Zip compressed data