Fix removing inferiors from within "thread apply" commands
authorPedro Alves <palves@redhat.com>
Wed, 19 Apr 2017 12:12:23 +0000 (13:12 +0100)
committerPedro Alves <palves@redhat.com>
Wed, 19 Apr 2017 12:12:23 +0000 (13:12 +0100)
commit3a3fd0fd2c4c87fdd588c51d879961a49e38f0c1
treec76d8c677a430fc2511e380fa895258e78811882
parent9bcb1f1630b05594fa86bfd017639cfcc966b11c
Fix removing inferiors from within "thread apply" commands

This patch fixes an internal error exposed by a test that does
something like:

  define kill-and-remove
    kill inferiors 2
    remove-inferiors 2
  end

  # Start one inferior.
  start

  # Start another inferior.
  add-inferior 2
  inferior 2
  start

  # Kill and remove inferior 1 while inferior 2 is selected.
  thread apply 1.1 kill-and-remove

The internal error looks like this:

 Thread 1.1 (Thread 0x7ffff7fc2700 (LWP 20677)):
 [Switching to inferior 1 [process 20677] (gdb/testsuite/outputs/gdb.threads/threadapply/threadapply)]
 [Switching to thread 1.1 (Thread 0x7ffff7fc2700 (LWP 20677))]
 #0  main () at src/gdb/testsuite/gdb.threads/threadapply.c:38
 38          for (i = 0; i < NUM; i++)
 src/gdb/inferior.c:66: internal-error: void set_current_inferior(inferior*): Assertion `inf != NULL' failed.
 A problem internal to GDB has been detected,
 further debugging may prove unreliable.
 Quit this debugging session? (y or n) FAIL: gdb.threads/threadapply.exp: kill_and_remove_inferior: try kill-and-remove: thread apply 1.1 kill-and-remove (GDB internal error)

There are several problems around this area of the code.  One is that
in do_restore_current_thread_cleanup, we do a look up of inferior by
ptid, which can find the wrong inferior if the previously selected
inferior exited and some other inferior was started with a reused pid
(rare, but still...).

The other problem is that the "remove-inferiors" command rejects
attempts to remove the current inferior, but when we get to
"remove-inferiors" in a "thread apply THR remove-inferiors 2" command,
the current inferior is the inferior of thread THR, not the previously
selected inferior, so if the previously selected inferior was inferior
2, that command still manages to wipe it, and then gdb restores the
old selected inferior, which is now a dangling pointer...

So the fix here is:

- Make make_cleanup_restore_current_thread store a pointer to the
  previously selected inferior directly, and use it directly instead
  of doing ptid look ups.

- Add a refcount to inferiors, very similar to thread_info's refcount,
  that is incremented/decremented by
  make_cleanup_restore_current_thread, and checked before deleting an
  inferior.  To avoid duplication, a new refcounted_object type is
  added, that both thread_info and inferior inherit from.

gdb/ChangeLog:
2017-04-19  Pedro Alves  <palves@redhat.com>

* common/refcounted-object.h: New file.
* gdbthread.h: Include "common/refcounted-object.h".
(thread_info): Inherit from refcounted_object and add comments.
(thread_info::incref, thread_info::decref)
(thread_info::m_refcount): Delete.
(thread_info::deletable): Use the refcounted_object::refcount()
method.
* inferior.c (current_inferior_): Add comment.
(set_current_inferior): Increment/decrement refcounts.
(prune_inferiors, remove_inferior_command): Skip inferiors marked
not-deletable instead of comparing with the current inferior.
(initialize_inferiors): Increment the initial inferior's refcount.
* inferior.h (struct inferior): Forward declare.
Include "common/refcounted-object.h".
(current_inferior, set_current_inferior): Move declaration to
before struct inferior's definition, and fix comment.
(inferior): Inherit from refcounted_object.  Add comments.
* thread.c (switch_to_thread_no_regs): Reference the thread's
inferior pointer directly instead of doing a ptid lookup.
(switch_to_no_thread): New function.
(switch_to_thread(thread_info *)): New function, factored out
from ...
(switch_to_thread(ptid_t)): ... this.
(restore_current_thread): Delete.
(current_thread_cleanup): Remove 'inf_id' and 'was_removable'
fields, and add 'inf' field.
(do_restore_current_thread_cleanup): Check whether old->inf is
alive instead of looking up an inferior by ptid.  Use
switch_to_thread and switch_to_no_thread.
(restore_current_thread_cleanup_dtor): Use old->inf directly
instead of lookup up an inferior by id.  Decref the inferior.
Don't restore 'removable'.
(make_cleanup_restore_current_thread): Same the inferior pointer
in old, instead of the inferior number.  Incref the inferior.
Don't save/clear 'removable'.

gdb/testsuite/ChangeLog:
2017-04-19  Pedro Alves  <palves@redhat.com>

* gdb.threads/threadapply.exp (kill_and_remove_inferior): New
procedure.
(top level): Call it.
* lib/gdb.exp (gdb_define_cmd): New procedure.
gdb/ChangeLog
gdb/common/refcounted-object.h [new file with mode: 0644]
gdb/gdbthread.h
gdb/inferior.c
gdb/inferior.h
gdb/testsuite/ChangeLog
gdb/testsuite/gdb.threads/threadapply.exp
gdb/testsuite/lib/gdb.exp
gdb/thread.c