PR gdb/20609 - attach of JIT-debug-enabled inf 7.11.1 regression
authorJan Kratochvil <jan.kratochvil@redhat.com>
Thu, 29 Sep 2016 15:38:16 +0000 (17:38 +0200)
committerJan Kratochvil <jan.kratochvil@redhat.com>
Thu, 29 Sep 2016 15:39:39 +0000 (17:39 +0200)
commitbb805577d2b212411fb7b0a2d01644567fac4e8d
tree975c7465c2f346d1122bc64221db7ca2ed176c9c
parent50751e18f3f2fc47959a577a5754f1a2d80baf18
PR gdb/20609 - attach of JIT-debug-enabled inf 7.11.1 regression

Regression: gdb --pid $(pidof qemu-system-x86_64) stopped working with gdb 7.11.1
https://sourceware.org/bugzilla/show_bug.cgi?id=20609

It was reported for qemu-system-x86_64 but it happens for any multithreaded
inferior with a JIT debugging hook.

136613ef0c6850427317e57be1b644080ff6decb is the first bad commit
Author: Pedro Alves <palves@redhat.com>
    Fix PR gdb/19828: gdb -p <process from a container>: internal error
Message-ID: <cbdf2e04-4fa8-872a-2a23-08c9c1b26e00@redhat.com>
https://sourceware.org/ml/gdb-patches/2016-05/msg00450.html

jit_breakpoint_re_set() is specific by trying to insert a breakpoint into the
main executable, not into a shared library.  During attachment GDB thinks it
needs to use 'breakpoint always-inserted' from
breakpoints_should_be_inserted_now() as a newly attached thread is
'thread_info->executing' due to 'lwp_info->must_set_ptrace_flags' enabled and
the task not yet stopped.  This did not happen before the 'bad commit' above
which adds tracking of such thread.

GDB then fails to insert the breakpoints to invalid address as PIE executable
gets properly relocated during later phase of attachment.  One can see in the
backtraces below:
 -> jit_breakpoint_re_set_internal()
later:
 -> svr4_exec_displacement()

One can suppress the initial breakpoint_re_set() call as there will be another
breakpoint_re_set() done from the final post_create_inferior() call in
setup_inferior().

BTW additionally 'threads_executing' cache bool is somehow stale (somewhere is
missing update_threads_executing()).  I was trying to deal with that in my
first/second attempt below but in my final third attempt (attached) I have
left it as it is.

First attempt trying not to falsely require 'breakpoint always-inserted':
  https://people.redhat.com/jkratoch/rhbz1375553-fix1.patch
Reduced first attempt:
  https://people.redhat.com/jkratoch/rhbz1375553-fix2.patch

The third attempt suppresses breakpoint insertion until PIE executable gets
relocated by svr4_exec_displacement().  Applied.

gdb/ChangeLog
2016-09-29  Jan Kratochvil  <jan.kratochvil@redhat.com>

PR gdb/20609 - attach of JIT-debug-enabled inf 7.11.1 regression
* exec.c (exec_file_locate_attach): Add parameter defer_bp_reset.
Use it.
* gdbcore.h (exec_file_locate_attach): Add parameter defer_bp_reset.
* infcmd.c (setup_inferior): Update caller.
* remote.c (remote_add_inferior): Likewise.

gdb/testsuite/ChangeLog
2016-09-29  Jan Kratochvil  <jan.kratochvil@redhat.com>

PR gdb/20609 - attach of JIT-debug-enabled inf 7.11.1 regression
* gdb.base/jit-attach-pie.c: New file.
* gdb.base/jit-attach-pie.exp: New file.
gdb/ChangeLog
gdb/exec.c
gdb/gdbcore.h
gdb/infcmd.c
gdb/remote.c
gdb/testsuite/ChangeLog
gdb/testsuite/gdb.base/jit-attach-pie.c [new file with mode: 0644]
gdb/testsuite/gdb.base/jit-attach-pie.exp [new file with mode: 0644]