Re: [RFC] mask off is-syscall bit for TRAP_IS_SYSCALL

[Pedro Alves <pedro@codesourcery.com>]

A Monday 28 September 2009 20:58:48, Pedro Alves escreveu:
> On Monday 28 September 2009 20:39:05, Doug Evans wrote:
> > Hi.
> > 
> > On one system I use (bi-arch ubuntu-hardy clone),
> > i386-disp-step.exp is failing because the wait status
> > value linux_nat_wait_1 gets when hitting a system call is 0857f
> > which gets passed to the upper layers which then get confused
> > by a signal of 0x85 (== 0x80 | SIGTRAP).
> > 
> > This patch fixes things by masking off the 0x80 bit before
> > passing the signal number to up the call chain.
> > 
> > Ok to check in?
> 
> This seems OK-is to me, although I see one extra case that
> isn't handled correctly:
> 
> stop_wait_callback:
> 
>       status = wait_lwp (lp);
> 
> ...
> 
>       if (WSTOPSIG (status) != SIGSTOP)
> 	{
> 	  if (WSTOPSIG (status) == SIGTRAP)
> 	    {
> ...
>             }
>           else
>             {
> 	      /* If the lp->status field is still empty, use it to
> 		 hold this event.  If not, then this event must be
> 		 returned to the event queue of the LWP.  */
> 	      if (lp->status)
> 		{
> 		  if (debug_linux_nat)
> 		    {
> 		      fprintf_unfiltered (gdb_stdlog,
> 					  "SWC: kill %s, %s\n",
> 					  target_pid_to_str (lp->ptid),
> 					  status_to_str ((int) status));
> 		    }
> 		  kill_lwp (GET_LWP (lp->ptid), WSTOPSIG (status));      <<<<<<<<
> 		}
> 
> It seems we can reach that <<< marked code with a TRAP_IS_SYSCALL, but,
> I doubt that we want to requeue that signal in the kernel (?).
> 

In case I wasn't clear, I was alluding at the fact that I think
the filtering is done too late..  In fact, things are worse than I
imagined.   See here for a simple failing example:

#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <pthread.h>

void *
thread_function (void *arg)
{
  while (1)
    {
      usleep (1);
    }
}

int
main ()
{
  int res;
  pthread_t thread;
  long i = 0;

  for (i = 0; i < 10; i++)
    {
      res = pthread_create(&thread,
			   NULL,
			   thread_function,
			   NULL);
    }

  thread_function (NULL);
}


 (gdb) r
 Starting program: /home/pedro/gdb/tests/trap_is_syscall
 [New Thread 0x7fe7dee066e0 (LWP 9531)]
 [Thread debugging using libthread_db enabled]
 [New Thread 0x40800950 (LWP 9538)]
 [New Thread 0x41001950 (LWP 9539)]
 [New Thread 0x41802950 (LWP 9540)]
 [New Thread 0x42003950 (LWP 9541)]
 [New Thread 0x42804950 (LWP 9542)]
 [New Thread 0x43005950 (LWP 9545)]
 [New Thread 0x43806950 (LWP 9547)]
 [New Thread 0x44007950 (LWP 9548)]
 [New Thread 0x44808950 (LWP 9549)]
 [New Thread 0x45009950 (LWP 9550)]

<ctrl-c>

 Program received signal SIGINT, Interrupt.
 0x00007ffff78ffb81 in nanosleep () from /lib/libc.so.6
 (gdb) catch syscall
 warning: Could not open "syscalls/amd64-linux.xml"
 warning: Could not load the syscall XML file `syscalls/amd64-linux.xml'.
 GDB will not be able to display syscall names.
 Catchpoint 1 (any syscall)
 (gdb) c
 Continuing.
 [Switching to Thread 0x45009950 (LWP 9550)]

 Catchpoint 1 (call to syscall 35), 0x00007ffff78ffb81 in nanosleep () from /lib/libc.so.6
 (gdb) c
 Continuing.

 Program received signal ?, Unknown signal.
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 [Switching to Thread 0x44808950 (LWP 9549)]
 0x00007ffff78ffb81 in nanosleep () from /lib/libc.so.6
 (gdb)      

`?' is due to TRAP_IS_SYSCALL.

Same with "set debug lin-lwp 1":

 linux_nat_wait: [process -1]
 LLW: Using pending wait status Trace/breakpoint trap (stopped at syscall) for Thread 0x44808950 (LWP 10484).
 LLW: Candidate event Trace/breakpoint trap (stopped at syscall) in Thread 0x44808950 (LWP 10484).

 Program received signal ?, Unknown signal.
 [Switching to Thread 0x44808950 (LWP 10484)]
 0x00007ffff78ffb81 in nanosleep () from /lib/libc.so.6

I wouldn't be surprised if syscall events were inverted from here
on (entry/exit).


Playing with a patch like the below, it's easier to trigger the case
I was mentioning before:

---
 gdb/linux-nat.c |   11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

Index: src/gdb/linux-nat.c
===================================================================
--- src.orig/gdb/linux-nat.c	2009-09-29 00:09:13.000000000 +0100
+++ src/gdb/linux-nat.c	2009-09-29 00:13:59.000000000 +0100
@@ -2400,6 +2400,13 @@ stop_wait_callback (struct lwp_info *lp,
 				    target_pid_to_str (lp->ptid),
 				    errno ? safe_strerror (errno) : "OK");
 
+	      if (WSTOPSIG (status) == TRAP_IS_SYSCALL)
+		{
+		  /* Simulate the case of a signal < SIGTRAP and
+		     SIGSTOP being delivered before the SIGSTOP.  */
+		  // kill_lwp (GET_LWP (lp->ptid), SIGINT);
+		}
+
 	      /* Hold this event/waitstatus while we check to see if
 	         there are any more (we still want to get that SIGSTOP). */
 	      stop_wait_callback (lp, NULL);
@@ -2416,7 +2423,9 @@ stop_wait_callback (struct lwp_info *lp,
 					  target_pid_to_str (lp->ptid),
 					  status_to_str ((int) status));
 		    }
-		  kill_lwp (GET_LWP (lp->ptid), WSTOPSIG (status));
+
+		  // gdb_assert (WSTOPSIG (status) != TRAP_IS_SYSCALL);
+		  gdb_assert (kill_lwp (GET_LWP (lp->ptid), WSTOPSIG (status)) == 0);
 		}
 	      else
 		lp->status = status;

Uncommenting the first commented out kill_lwp the patchlet adds, makes
the last gdb_assert trigger (can't kill with signal 0x85), meaning, a syscall event
gets lost.

-- 
Pedro Alves