Re: Remote debugging of multithreaded programs using gdbserver

[Daniel Jacobowitz <drow@mvista.com>]

First of all, gdbserver has been essentially rewritten between 5.1.1 and
5.2.  I recommend you look at the current CVS.  I have a great deal of
(thread-related) work pending, also.

Second - do you have details on the "broken" death event reporting in
glibc 2.1.3?  Does it make the death events unreliable, or outright
harmful?

On Thu, Apr 18, 2002 at 10:29:21AM +0530, Subhashini Nagarajan Rao wrote:
>   Since event reporting facility is broken for TD_DEATH events in glibc 2.1.3
>   and the present implementation uses this version of glibc, thread deaths are
>   are handled differently. While an immediate reporting of thread creation
>   to gdbserver is mandatory to ensure that the gdbserver traces the newly
>   created threads with immediate effect, thread death updation at the client
>   needs could be procrastinated to the time when one of the following
>   events takes place:
>   (i) info threads command is fired at the client by the user.
>   (ii) context switch happens from one thread to another during execution.

"context switch" is misleading.  The context switches of the process
are irrelevant.  I assume you mean whenever the client GDB requests a
change of thread...

>   Gdb Client related and protocol enhancements towards the same are:
>   
>   The client sends a qSymbol:: packet to notify the target that
>   it is prepared to service symbol look-up requests, typically when new
>   objects are loaded. On the initial receipt of a qSymbol:: ,
>   the server enables thread event reporting and sends the thread
>   creation notification address through qSymbol! packet for the client to add
>   breakpoint at that address. qSymbol! packet essentially sticks to the the
>   following format:
>   
>   qSymbol!<thread creation notification address>:<thread death notification address>

I don't think this is the right approach.  The client GDB should not be
aware of anything involving thread_db.  It's none of its business.

OTOH, the bugs which cause thread_db to be pushed on the target stack
when connected to a remote target definitely need to be fixed.  I'm
not quite decided on how to do this yet.

Why do you bother with this if you have obviously given the server the
ability to place breakpoints?

>   Handling of the Protocol packets :
>   
>   *) qfThreadInfo : When this packet is received libthread's thread
>   iterator function is called to collect all the active thread
>   ids. The collected thread ids are maintained in a
>   linked list, thread_list. These thread ids are then
>   copied from this list to the buffer to be sent to the
>   client. The purpose behind storing the thread ids
>   rather than writing directly into the buffer is to
>   overcome the limitation posed by the buffer size.
>   
>   *) qsThreadInfo : Copy from the next thread id from the thread_list
>   to the buffer if any more are to be copied else send 'l' as the reply. 
>   Also free the thread_list once done.

This information should be cached, there's no point in looking it up
from scratch every time you are queried.

>   4. Adding breakpoints in the threads.
>   
>   The list of breakpoints in the program is maintained as a linked list
>   of breakpoint addresses. The breakpoint list is a doubly linked list
>   having the breakpoint address and a 2 character array('cc' for IA32).
>   A free pool of these nodes are maintained from where the breakpoint list
>   draws out nodes as and when the need arises.

"A two character array"?  0xCC is the code for a one-byte i386
breakpoint, but I can't imagine what else you mean by this. 
Breakpoints can be a word, or possibly more.

>   When one thread is stopped all the other threads are also stopped so that all
>   are in sync and in control under the debugger. This is
>   done by calling stop_all_threads(). If the thread stopped is single-stepped
>   then preference is given to this thread and allowed to resume next. 
>   When none of the threads are single-stepped and more than 1 
>   thread has received a SIGTRAP, randomnly 1 thread is chosen among the threads
>   which received the trap signal and allowed to resume next 
>   If any of the other threads has got a SIGTRAP because of a gdb 
>   inserted breakpoint, the pc values have to be re-adjusted to ensure that 
>   these threads re-execute the breakpoint instruction. The selection of the 
>   thread to resume next is handled by select_event_pid(). 

All of this appears to be taken straight from GDB's native thread debug
package.  None of it is clearly thought out in this context, IMHO. 
Limitations in GDB make it hard to fix these properly there (I spent a
full day trying recently...) but gdbserver has plenty of room.  For
instance, I'd wager you stop all threads on a thread creation event. 
Also, there's no reason to do the fancy PC-adjustment.  You should be
able to save that stop event and re-report it later if nothing disturbs
it.

>   8. Miscellaneous changes
>   
>   When the user fires a 'thread <id>' command , switch to the specified
>   thread but a step after this was not able to tell the server that the
>   thread to be stepped is the new thread and not the previous one. Hence,
>   there is a need for the client, to remotely specify the current thread
>   in switch_to_thread(). However since switch_to_thread() is written

I don't see anything here that requires information gdbserver does not
already have.  It knows the last thread stepped and the thread
requested.

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer