gdb: disable commit resumed in target_kill

New in this version:

 - Better comment in target_kill
 - Uncomment line in test to avoid hanging when exiting, when testing on
   native-extended-gdbserver

PR 28275 shows that doing a sequence of:

 - Run inferior in background (run &)
 - kill that inferior
 - Run again

We get into this assertion:

    /home/smarchi/src/binutils-gdb/gdb/target.c:2590: internal-error: target_wait: Assertion `!proc_target->commit_resumed_state' failed.

    #0  internal_error_loc (file=0x5606b344e740 "/home/smarchi/src/binutils-gdb/gdb/target.c", line=2590, fmt=0x5606b344d6a0 "%s: Assertion `%s' failed.") at /home/smarchi/src/binutils-gdb/gdbsupport/errors.cc:54
    #1  0x00005606b6296475 in target_wait (ptid=..., status=0x7fffb9390630, options=...) at /home/smarchi/src/binutils-gdb/gdb/target.c:2590
    #2  0x00005606b5767a98 in startup_inferior (proc_target=0x5606bfccb2a0 <the_amd64_linux_nat_target>, pid=3884857, ntraps=1, last_waitstatus=0x0, last_ptid=0x0) at /home/smarchi/src/binutils-gdb/gdb/nat/fork-inferior.c:482
    #3  0x00005606b4e6c9c5 in gdb_startup_inferior (pid=3884857, num_traps=1) at /home/smarchi/src/binutils-gdb/gdb/fork-child.c:132
    #4  0x00005606b50f14a5 in inf_ptrace_target::create_inferior (this=0x5606bfccb2a0 <the_amd64_linux_nat_target>, exec_file=0x604000039f50 "/home/smarchi/build/binutils-gdb/gdb/test", allargs="", env=0x61500000a580, from_tty=1)
        at /home/smarchi/src/binutils-gdb/gdb/inf-ptrace.c:105
    #5  0x00005606b53b6d23 in linux_nat_target::create_inferior (this=0x5606bfccb2a0 <the_amd64_linux_nat_target>, exec_file=0x604000039f50 "/home/smarchi/build/binutils-gdb/gdb/test", allargs="", env=0x61500000a580, from_tty=1)
        at /home/smarchi/src/binutils-gdb/gdb/linux-nat.c:978
    #6  0x00005606b512b79b in run_command_1 (args=0x0, from_tty=1, run_how=RUN_NORMAL) at /home/smarchi/src/binutils-gdb/gdb/infcmd.c:468
    #7  0x00005606b512c236 in run_command (args=0x0, from_tty=1) at /home/smarchi/src/binutils-gdb/gdb/infcmd.c:526

When running the kill command, commit_resumed_state for the
process_stratum_target (linux-nat, here) is true.  After the kill, when
there are no more threads, commit_resumed_state is still true, as
nothing touches this flag during the kill operation.  During the
subsequent run command, run_command_1 does:

    scoped_disable_commit_resumed disable_commit_resumed ("running");

We would think that this would clear the commit_resumed_state flag of
our native target, but that's not the case, because
scoped_disable_commit_resumed iterates on non-exited inferiors in order
to find active process targets.  And after the kill, the inferior is
exited, and the native target was unpushed from it anyway.  So
scoped_disable_commit_resumed doesn't touch the commit_resumed_state
flag of the native target, it stays true.  When reaching target_wait, in
startup_inferior (to consume the initial expect stop events while the
inferior is starting up and working its way through the shell),
commit_resumed_state is true, breaking the contract saying that
commit_resumed_state is always false when calling the targets' wait
method.

(note: to be correct, I think that startup_inferior should toggle
commit_resumed between the target_wait and target_resume calls, but I'll
ignore that for now)

I can see multiple ways to fix this.  In the end, we need
commit_resumed_state to be cleared by the time we get to that
target_wait.  It could be done at the end of the kill command, or at the
beginning of the run command.

To keep things in a coherent state, I'd like to make it so that after
the kill command, when the target is left with no threads, its
commit_resumed_state flag is left to false.  This way, we can keep
working with the assumption that a target with no threads (and therefore
no running threads) has commit_resumed_state == false.

Do this by adding a scoped_disable_commit_resumed in target_kill.  It
clears the target's commit_resumed_state on entry, and leaves it false
if the target does not have any resumed thread on exit.  That means,
even if the target has another inferior with stopped threads,
commit_resumed_state will be left to false, which makes sense.

Add a test that tries to cover various combinations of actions done
while an inferior is running (and therefore while commit_resumed_state
is true on the process target).

Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=28275
Change-Id: I8e6fe6dc1f475055921520e58cab68024039a1e9
Approved-By: Andrew Burgess <aburgess@redhat.com>

1 files changed, 9 insertions, 0 deletions

diff --git a/gdb/target.c b/gdb/target.c
index 0c86b57..0eae530 100644
--- a/gdb/target.c
+++ b/gdb/target.c
@@ -908,6 +908,15 @@ add_deprecated_target_alias (const target_info &tinfo, const char *alias)
 void
 target_kill (void)
 {
+
+  /* If the commit_resume_state of the to-be-killed-inferior's process stratum
+     is true, and this inferior is the last live inferior with resumed threads
+     of that target, then we want to leave commit_resume_state to false, as the
+     target won't have any resumed threads anymore.  We achieve this with
+     this scoped_disable_commit_resumed.  On construction, it will set the flag
+     to false.  On destruction, it will only set it to true if there are resumed
+     threads left.  */
+  scoped_disable_commit_resumed disable ("killing");
   current_inferior ()->top_target ()->kill ();
 }