RFA: nptl threading support patches

["J. Johnston" <jjohnstn at redhat dot com>]

The attached patch adds nptl threading support for gdb.

There are a couple of key design changes to note with the nptl-enabled
kernel.  First of all, kill semantics have changed.  It is no longer
possible to send a signal to an lwp via kill.  Only a process id can be
used, rather than a thread lwp.  To keep the lin-lwp logic whereby SIGSTOP
is applied to a particular thread and some signals are pushed back, the
tkill syscall must be used.  The tkill syscall allows us to send a signal
to a particular thread.  A configuration check is made to see if we can
call syscall(__NR_tkill,..) and a runtime check is made to ensure that we can call
it successfully without getting ENOSYS.

Another key behavior change regards thread exit.  For linuxthreads, exiting
causes a WIFEXITED event to occur at which point we can delete the current
thread and determine if any threads still exist.  In the new nptl model,
only the main thread generates this event and does so only once all the
threads have exited.  There is no event for the individual threads exiting
so we must check at various times to see if a thread has gone away.  For
example, when we stop all of the threads, this is a good time to see if
any have gone away.  When we get the exit signal, we check if we have the
main thread or not.  If it is the main thread, we have to stop all threads
(which will check whether they are alive or not), and then see how many
threads are left.  This logic works for either model.  If there are still
threads alive, we resume them and continue on without reporting the
exit.

A change was made in thread-db.c to handle a FIXME that has been
there for a while.  When we get a create breakpoint in check_event, we
do not know what lwp has been created so we call td_ta_event_getmsg() to
get "a message" off the queue.  This message might not actually be the
message corresponding to the breakpoint that stopped us.  I have changed
the logic to perform a loop and read off all messages on the queue.  This
enables death event reporting to work if desired.  Without it, you can
get scenarios such as a death event preceding its create event, etc..

The testsuite needed some changes because there is no longer a manager
thread with nptl so the testcase must account for either model when
doing info threads.  The schedlock testcase was changed because
nptl scheduling differs from linuxthreads.  One can't assume that
all threads will run equally in a small time slice, most notably when
stepping a particular thread.

It is a lot to digest.  Let me know if/when it is ok to commit.

-- Jeff J.

2003-04-17  Jeff Johnston  <jjohnstn at redhat dot com>

	* acconfig.h: Add HAVE_TKILL_SYSCALL definition check.
	* config.in: HAVE_TKILL_SYSCALL and HAVE_SYSCALL checks added.
	* configure.in: Add test for syscall function and check for
	__NR_tkill macro in <syscall.h> to set HAVE_TKILL_SYSCALL.
	* configure: Regenerated.
	* thread-db.c (check_event): For create/death event breakpoints,
	loop through all messages to ensure that we read the message
	corresponding to the breakpoint we are at.
	* lin-lwp.c [HAVE_TKILL_SYSCALL]: Include <unistd.h> and <sys/syscall.h>.
	(kill_lwp): New function that uses tkill syscall or
	uses kill, depending on whether threading model is nptl or not.
	All callers of kill() changed to use kill_lwp().
	(lin_lwp_wait): Make special check when WIFEXITED occurs to
	see if all threads have already exited in the nptl model.
	(stop_wait_callback): Check for threads already having exited and
	delete such threads fromt the lwp list when discovered.
	(stop_callback): Don't assert retcode of kill call.
	* i386-linux-nat.c (ps_get_thread_area): New function needed by
	nptl libthread_db.
	
2003-04-17  Jeff Johnston  <jjohnstn at redhat dot com>

	* schedlock.exp: Remove assumptions that all threads will run in a particular
	time slice.  Also ensure we break in one of the child threads rather than the
	main thread.

	Roland McGrath  <roland at redhat dot com>
	* linux-dp.exp: Account for fact that in nptl model there is no manager thread.

Attachment: nptl2.patch.tgz
Description: GNU Zip compressed data