The buildbots show that attach-many-short-lived-thread.exp is racy.
But after staring at debug logs and playing with SystemTap scripts for
a (long) while, I figured out that neither GDB, nor the kernel nor the
test's program itself are at fault.
The problem is simply that the testsuite machinery is currently
subject to PID-reuse races. The attach-many-short-lived-threads.c
test program just happens to be much more susceptible to trigger this
race because threads and processes share the same number space on
Linux, and the test spawns many many short lived threads in
succession, thus enlarging the race window a lot.
Part of the problem is that several tests spawn processes with "exec&"
(in order to test the "attach" command) , and then at the end of the
test, to make sure things are cleaned up, issue a 'remote_spawn "kill
-p $testpid"'. Since with tcl's "exec&", tcl itself is responsible
for reaping the process's exit status, when we go kill the process,
testpid may have already exited _and_ its status may have (and often
has) been reaped already. Thus it can happen that another process
meanwhile reuses $testpid, and that "kill" command kills the wrong
process... Frequently, that happens to be
attach-many-short-lived-thread, but this explains other test's races
as well.
In the attach-many-short-lived-threads test, it sometimes manifests
like this:
(gdb) file /home/pedro/gdb/mygit/build/gdb/testsuite/gdb.threads/attach-many-short-lived-threads
Reading symbols from /home/pedro/gdb/mygit/build/gdb/testsuite/gdb.threads/attach-many-short-lived-threads...done.
(gdb) Loaded /home/pedro/gdb/mygit/build/gdb/testsuite/gdb.threads/attach-many-short-lived-threads into /home/pedro/gdb/mygit/build/gdb/testsuite/../../gdb/gdb
attach 5940
Attaching to program: /home/pedro/gdb/mygit/build/gdb/testsuite/gdb.threads/attach-many-short-lived-threads, process 5940
warning: process 5940 is a zombie - the process has already terminated
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ptrace: Operation not permitted.
(gdb) PASS: gdb.threads/attach-many-short-lived-threads.exp: iter 1: attach
info threads
No threads.
(gdb) PASS: gdb.threads/attach-many-short-lived-threads.exp: iter 1: no new threads
set breakpoint always-inserted on
(gdb) PASS: gdb.threads/attach-many-short-lived-threads.exp: iter 1: set breakpoint always-inserted on
Other times the process dies while the test is ongoing (the process is
ptrace-stopped):
(gdb) print again = 1
Cannot access memory at address 0x6020cc
(gdb) FAIL: gdb.threads/attach-many-short-lived-threads.exp: iter 2: reset timer in the inferior
(Recall that on Linux, SIGKILL is not interceptable)
And other times it dies just while we're detaching:
$4 = 319
(gdb) PASS: gdb.threads/attach-many-short-lived-threads.exp: iter 2: print seconds_left
detach
Can't detach Thread 0x7fb13b7de700 (LWP 1842): No such process
(gdb) FAIL: gdb.threads/attach-many-short-lived-threads.exp: iter 2: detach
GDB mishandles the latter (it should ignore ESRCH while detaching just
like when continuing), but that's another story.
The fix here is to change spawn_wait_for_attach to use Expect's
'spawn' command instead of Tcl's 'exec&' to spawn programs, because
with spawn we control when to wait for/reap the process. That allows
killing the process by PID without being subject to pid-reuse races,
because even if the process is already dead, the kernel won't reuse
the process's PID until the zombie is reaped.
The other part of the problem lies in DejaGnu itself, unfortunately.
I have occasionally seen tests (attach-many-short-lived-threads
included, but not only that one) die with a random inexplicable
SIGTERM too, and that too is caused by the same reason, except that in
that case, the rogue SIGTERM is sent from this bit in DejaGnu's remote.exp:
exec sh -c "exec > /dev/null 2>&1 && (kill -2 $pgid || kill -2 $pid) && sleep 5 && (kill $pgid || kill $pid) && sleep 5 && (kill -9 $pgid || kill -9 $pid) &"
...
catch "wait -i $shell_id"
Even if the program exits promptly, that whole cascade of kills
carries on in the background, thus potentially killing the poor
process that manages to reuse $pid...
I sent a fix for that to the DejaGnu list:
http://lists.gnu.org/archive/html/dejagnu/2015-07/msg00000.html
With both patches in place, I haven't seen
attach-many-short-lived-threads.exp fail again.
Tested on x86_64 Fedora 20, native, gdbserver and extended-gdbserver.
gdb/testsuite/ChangeLog:
2015-07-31 Pedro Alves <palves@redhat.com>
* gdb.base/attach-pie-misread.exp: Rename $res to $test_spawn_id.
Use spawn_id_get_pid. Wait for spawn id after eof. Use
kill_wait_spawned_process instead of explicit "kill -9".
* gdb.base/attach-pie-noexec.exp: Adjust to spawn_wait_for_attach
returning a spawn id instead of a pid. Use spawn_id_get_pid and
kill_wait_spawned_process.
* gdb.base/attach-twice.exp: Likewise.
* gdb.base/attach.exp: Likewise.
(do_command_attach_tests): Use gdb_spawn_with_cmdline_opts and
gdb_test_multiple.
* gdb.base/solib-overlap.exp: Adjust to spawn_wait_for_attach
returning a spawn id instead of a pid. Use spawn_id_get_pid and
kill_wait_spawned_process.
* gdb.base/valgrind-infcall.exp: Likewise.
* gdb.multi/multi-attach.exp: Likewise.
* gdb.python/py-prompt.exp: Likewise.
* gdb.python/py-sync-interp.exp: Likewise.
* gdb.server/ext-attach.exp: Likewise.
* gdb.threads/attach-into-signal.exp (corefunc): Use
spawn_wait_for_attach, spawn_id_get_pid and
kill_wait_spawned_process.
* gdb.threads/attach-many-short-lived-threads.exp: Adjust to
spawn_wait_for_attach returning a spawn id instead of a pid. Use
spawn_id_get_pid and kill_wait_spawned_process.
* gdb.threads/attach-stopped.exp (corefunc): Use
spawn_wait_for_attach, spawn_id_get_pid and
kill_wait_spawned_process.
* gdb.base/break-interp.exp: Rename $res to $test_spawn_id.
Use spawn_id_get_pid. Wait for spawn id after eof. Use
kill_wait_spawned_process instead of explicit "kill -9".
* lib/gdb.exp (can_spawn_for_attach): Adjust comment.
(kill_wait_spawned_process, spawn_id_get_pid): New procedures.
(spawn_wait_for_attach): Use spawn instead of exec to spawn
processes. Don't map cygwin/windows pids here. Now returns a
spawn id list.
Hi,
While examining BuildBot's logs, I noticed:
<https://sourceware.org/ml/gdb-testers/2015-q3/msg03767.html>
gdb.threads/attach-into-signal.exp has two nested loops and don't use
unique messages. This commit fixes that. Pushed under the obvious
rule.
gdb/testsuite/ChangeLog:
2015-07-29 Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.threads/attach-into-signal.exp (corefunc): Use
with_test_prefix on nested loops, uniquefying the test messages.
As the result of the patch below, GDB updates thread list when a stop is
presented to user. The tests don't have to fetch thread list explicitly.
[PATCH 3/3] Fix non-stop regressions caused by "breakpoints always-inserted off" changes
https://sourceware.org/ml/gdb-patches/2014-09/msg00734.html
This patch is to remove the test code updating thread list.
Run these three tests many times on arm-linux-gnueabi and x86-linux.
No regressions.
gdb/testsuite:
2014-10-11 Yao Qi <yao@codesourcery.com>
* gdb.threads/thread-find.exp: Don't execute command
"info threads".
* gdb.threads/attach-into-signal.exp (corefunc): Likewise.
* gdb.threads/linux-dp.exp: Don't check the condition
$threads_created equals to zero.
Two modifications:
1. The addition of 2013 to the copyright year range for every file;
2. The use of a single year range, instead of potentially multiple
year ranges, as approved by the FSF.
* gdb.threads/attach-into-signal.exp (corefunc): Don't enable
lin-lwp output. Set SIGALRM to stop. Adjust tests to not rely on
gdb's internal debug output. For the non-threaded case, look for
"Program received signal SIGLARM", for the threaded case, peek at
the thread's siginfo.
General cleanup, make output test messages unique, and build
different executable files for the non-threaded and threaded
cases.
* gdb.threads/attach-into-signal.exp (binfile, escapedbinfile):
Delete.
(executable_nothr, executable_thr): New globals.
(top level): Adjust to delete both executables.
(corefunc): New parameter $executable. Set $pf_prefix instead of
hand writing a prefix in tests. Issue a clean_restart and enable
lin-lwp debug output here.
(top level): Adjust. Use build_executable. Don't start gdb here,
and don't enable lin-lwp debug output here.
* gdb.threads/Makefile.in (EXECUTABLES): Adjust.
* gdb.base/async-shell.exp: Skip the testfile for use_gdb_stub.
* gdb.base/attach-pie-misread.exp: Likewise.
* gdb.base/break-interp.exp: Likewise.
* gdb.base/default.exp (attach, run "r" abbreviation, run): Skip these
testcases for use_gdb_stub.
* gdb.dwarf2/dw2-restore.exp: Fix the run for use_gdb_stub.
(continue): New testcase.
* gdb.python/py-events.exp: Skip the testfile for use_gdb_stub.
* gdb.threads/attach-into-signal.exp: Likewise.
* gdb.threads/attach-stopped.exp: Likewise.
* gdb.threads/attachstop-mt.exp: Likewise.
Pedro Alves <pedro@codesourcery.com>
Based on work by Jan Kratochvil <jan.kratochvil@redhat.com> and Jeff
Johnston <jjohnstn@redhat.com>.
* NEWS: Mention attach to stopped process fix.
* infcmd.c (detach_command, disconnect_command): Discard the thread
list.
* infrun.c (handle_inferior_event): Do not ignore non-SIGSTOP while
attaching. Use signal_stop_state.
(signal_stop_state): Check stop_soon.
* linux-nat.c (kill_lwp): Declare earlier.
(pid_is_stopped, linux_nat_post_attach_wait): New.
(lin_lwp_attach_lwp): Use linux_nat_post_attach_wait. Update
comments.
(linux_nat_attach): Use linux_nat_post_attach_wait.
(detach_callback, linux_nat_detach): Improve handling for signalled
processes.
(linux_nat_pid_to_str): Always print out the LWP ID if it differs
from the process ID.
* Makefile.in (infcmd.o): Update.
2008-05-01 Jan Kratochvil <jan.kratochvil@redhat.com>
Daniel Jacobowitz <dan@codesourcery.com>
* gdb.threads/attach-into-signal.c, gdb.threads/attach-into-signal.exp,
gdb.threads/attach-stopped.c, gdb.threads/attach-stopped.exp,
gdb.threads/attachstop-mt.c, gdb.threads/attachstop-mt.exp: New.
