This is the mail archive of the gdb-patches@sources.redhat.com mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFA] Fix internal error in wait_lwp (interrupted system call)


On Thu, May 12, 2005 at 09:06:38PM +0200, Ulrich Weigand wrote:
> Hello,
> 
> we've had reports from our JVM/JIT development group that for them,
> gdb 6.3 frequently fails with internal errors like:
> linux-nat.c:1152: internal-error: wait_lwp: Assertion `pid == GET_LWP (lp->ptid)' failed.
> 
> It turned out that this happens when a SIGCHLD arrives during
> execution of the waitpid call.  This causes the signal handler
> to be executed, and subsequently the system call returns with
> errno equal to EINTR.
> 
> Now, looking through the linux-nat.c file, it would appear that this
> type of problem has been addressed at various places in different
> ways.  In linux_handle_extended_wait, the waitpid call is wrapped
> into an explicit do { } while (ret == -1 && errno == EINTR) loop.
> In linux_test_for_tracefork, this very loop is abstracted into a
> my_waitpid routine.  In child_wait and linux_nat_wait, there are
> larger loops that will handle this situation as well.  Finally,
> in lin_lwp_attach_lwp, SIGCHLD is actually blocked during the
> execution of the waitpid call.
> 
> However, there remain some places where waitpid is called without
> any such precaution, and wait_lwp is one of these.  When debugging
> a process making very heavy use of threads, as the JVM, this can
> lead to the error shown above.
> 
> Now, as far as I can see, there is really *no* place where GDB
> actually *wants* a system call to be interrupted by the SIGCHLD
> signal handler.  Thus, I'd propose to fix the problem at its 
> root by simply installing the handler with the SA_RESTART flag,
> causing any interrupted system call to be automatically restarted.
> 
> The patch below does this, and fixes all problems for the JVM team.
> It also passes regression testing on s390-ibm-linux and s390x-ibm-linux.
> 
> OK to commit?

On the one hand, this is very clever.  On the other hand, it's not very
robust.  This is not the only signal that could arrive.  Shouldn't
wait_lwp be looping on EINTR anyway, probably by using my_waitpid
(which is a recent addition)?


-- 
Daniel Jacobowitz
CodeSourcery, LLC


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]