Fix crash if connection drops in scoped_restore_current_thread's ctor, part 2
authorPedro Alves <pedro@palves.net>
Thu, 9 Jul 2020 17:14:09 +0000 (18:14 +0100)
committerPedro Alves <palves@redhat.com>
Fri, 10 Jul 2020 22:55:44 +0000 (23:55 +0100)
commitcce20f107400557f5c6b917babe7ff76fb1ec86e
tree50f669eb895d044c52697bcd578c26297ef705fa
parent6d7aa59270373b6b1de6ac28e40ebf972028ee3e
Fix crash if connection drops in scoped_restore_current_thread's ctor, part 2

Running the testsuite against an Asan-enabled build of GDB makes
gdb.base/multi-target.exp expose this bug.

scoped_restore_current_thread's ctor calls get_frame_id to record the
selected frame's ID to restore later.  If the frame ID hasn't been
computed yet, it will be computed on the spot, and that will usually
require accessing the target's memory and registers.  If the remote
connection closes, while we're computing the frame ID, the remote
target exits its inferiors, unpushes itself, and throws a
TARGET_CLOSE_ERROR error.  Exiting the inferiors deletes the
inferior's threads.

scoped_restore_current_thread increments the current thread's refcount
to prevent the thread from being deleted from under its feet.
However, the code that does that isn't considering the case of the
thread being deleted from within get_frame_id.  It only increments the
refcount _after_ get_frame_id returns.  So if the current thread is
indeed deleted, the

     tp->incref ();

statement references a stale TP pointer.

Incrementing the refcounts earlier fixes it.

We should probably also let the TARGET_CLOSE_ERROR error propagate in
this case.  That alone would fix it, though it seems better to tweak
the refcount handling too.  And to avoid having to manually decref
before throwing, convert to use gdb::ref_ptr.

Unfortunately, we can't define inferior_ref in inferior.h and then use
it in scoped_restore_current_thread, because
scoped_restore_current_thread is defined before inferior is
(inferior.h includes gdbthread.h).  To break that dependency, we would
have to move scoped_restore_current_thread to its own header.  I'm not
doing that here.

gdb/ChangeLog:

* gdbthread.h (inferior_ref): Define.
(scoped_restore_current_thread) <m_thread>: Now a thread_info_ref.
(scoped_restore_current_thread) <m_inf>: Now an inferior_ref.
* thread.c
(scoped_restore_current_thread::restore):
Adjust to gdb::ref_ptr.
(scoped_restore_current_thread::~scoped_restore_current_thread):
Remove manual decref handling.
(scoped_restore_current_thread::scoped_restore_current_thread):
Adjust to use
inferior_ref::new_reference/thread_info_ref::new_reference.
Incref the thread before calling get_frame_id instead of after.
Let TARGET_CLOSE_ERROR propagate.
gdb/ChangeLog
gdb/gdbthread.h
gdb/thread.c