binutils-gdb/gdb/testsuite/gdb.threads/attach-slow-waitpid.exp
Andrew Burgess 7010835a6c gdb: Don't drop SIGSTOP during stop_all_threads
This patch fixes an issue where GDB would sometimes hang when
attaching to a multi-threaded process.  This issue was especially
likely to trigger if the machine (running the inferior) was under
load.

In summary, the problem is an imbalance between two functions in
linux-nat.c, stop_callback and stop_wait_callback.  In stop_callback
we send SIGSTOP to a thread, but _only_ if the thread is not already
stopped, and if it is not signalled, which means it should stop soon.
In stop_wait_callback we wait for the SIGSTOP to arrive, however, we
are aware that the thread might have been signalled for some other
reason, and so if a signal other than SIGSTOP causes the thread to
stop then we stash that signal away so it can be reported back later.
If we get a SIGSTOP then this is discarded, after all, this signal was
sent from stop_callback.  Except that this might not be the case, it
could be that SIGSTOP was sent to a thread from elsewhere in GDB, in
which case we would not have sent another SIGSTOP from stop_callback
and the SIGSTOP received in stop_wait_callback should not be ignored.

Below I've laid out the exact sequence of events that I saw that lead
me to track down the above diagnosis.

After attaching to the inferior GDB sends a SIGSTOP to all of the
threads and then returns to the event loop waiting for interesting
things to happen.

Eventually the first target event is detected (this will be the first
SIGSTOP arriving) and GDB calls inferior_event_handler which calls
fetch_inferior_event.  Inside fetch_inferior_event GDB calls
do_target_wait which calls target_wait to find a thread with an event.

The target_wait call ends up in linux_nat_wait_1, which first checks
to see if any threads already have stashed stop events to report, and
if there are none then we enter a loop fetching as many events as
possible out of the kernel.  This event fetching is non-blocking, and
we give up once the kernel has no more events ready to give us.

All of the events from the kernel are passed through
linux_nat_filter_event which stashes the wait status for all of the
threads that reported a SIGSTOP, these will be returned by future
calls to linux_nat_wait_1.

Lets assume for a moment that we've attached to a multi-threaded
inferior, and that all but one thread has reported its stop during the
initial wait call in linux_nat_wait_1.  The other thread will be
reporting a SIGSTOP, but the kernel has not yet managed to deliver
that signal to GDB before GDB gave up waiting and continued handling
the events it already had.  GDB selects one of the threads that has
reported a SIGSTOP and passes this thread ID back to
fetch_inferior_event.

To handle the thread's SIGSTOP, GDB calls handle_signal_stop, which
calls stop_all_threads, this calls wait_one, which in turn calls
target_wait.

The first call to target_wait at this point will result in a stashed
wait status being returned, at which point we call setup_inferior.
The call to setup_inferior leads to a call into try_thread_db_load_1
which results in a call to linux_stop_and_wait_all_lwps.  This in turn
calls stop_callback on each thread followed by stop_wait_callback on
each thread.

We're now ready to make the mistake.  In stop_callback we see that our
problem thread is not stopped, but is signalled, so it should stop
soon.  As a result we don't send another SIGSTOP.

We then enter stop_wait_callback, eventually the problem thread stops
with SIGSTOP which we _incorrectly_ assume came from stop_callback,
and we discard.

Once stop_wait_callback has done its damage we return from
linux_stop_and_wait_all_lwps, finish in try_thread_db_load_1, and
eventually unwind back to the call to setup_inferior in
stop_all_threads.  GDB now loops around, and performs another
target_wait to get the next event from the inferior.

The target_wait calls causes us to once again reach linux_nat_wait_1,
and we pass through some code that calls resume_stopped_resumed_lwps.
This allows GDB to resume threads that are physically stopped, but
which GDB doesn't see any good reason for the thread to remain
stopped.  In our case, the problem thread which had its SIGSTOP
discarded is stopped, but doesn't have a stashed wait status to
report, and so GDB sets the thread going again.

We are now stuck waiting for an event on the problem thread that might
never arrive.

When considering how to write a test for this bug I struggled.  The
issue was only spotted _randomly_ when a machine was heavily loaded
with many multi-threaded applications, and GDB was being attached (by
script) to all of these applications in parallel.  In one reproducer I
required around 5 applications each of 5 threads per machine core in
order to reproduce the bug 2 out of 3 times.

What we really want to do though is simulate the kernel being slow to
report events through waitpid during the initial attach.  The solution
I came up with was to write an LD_PRELOAD library that intercepts
(some) waitpid calls and rate limits them to one per-second.  Any more
than that simply return 0 indicating there's no event available.
Obviously this can only be applied to waitpid calls that have the
WNOHANG flag set.

Unfortunately, once you ignore a waitpid call GDB can get a bit stuck.
Usually, once the kernel has made a child status available to waitpid
GDB will be sent a SIGCHLD signal.  However, if the kernel makes 5
child statuses available but, due to the preload library we only
collect one of them, then the kernel will not send any further SIGCHLD
signals, and so, when GDB, thinking that the remaining statuses have
not yet arrived sits waiting for a SIGCHLD it will be disappointed.

The solution, implemented within the preload library, is that, when we
hold back a waitpid result from GDB we spawn a new thread.  This
thread delays for a short period, and then sends GDB a SIGCHLD.  This
causes GDB to retry the waitpid, at which point sufficient time has
passed and our library allows the waitpid call to complete.

gdb/ChangeLog:

	* linux-nat.c (stop_wait_callback): Don't discard SIGSTOP if it
	was requested by GDB.

gdb/testsuite/ChangeLog:

	* gdb.threads/attach-slow-waitpid.c: New file.
	* gdb.threads/attach-slow-waitpid.exp: New file.
	* gdb.threads/slow-waitpid.c: New file.
2018-06-16 01:03:57 +01:00

101 lines
3.3 KiB
Plaintext

# Copyright 2018 Free Software Foundation, Inc.
#
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
# This test script tries to expose a bug in some of the uses of
# waitpid in the Linux native support within GDB. The problem was
# spotted on systems which were heavily loaded when attaching to
# threaded test programs. What happened was that during the initial
# attach, the loop of waitpid calls that normally received the stop
# events from each of the threads in the inferior was not receiving a
# stop event for some threads (the kernel just hadn't sent the stop
# event yet).
#
# GDB would then trigger a call to stop_all_threads which would
# continue to wait for all of the outstanding threads to stop, when
# the outstanding stop events finally arrived GDB would then
# (incorrectly) discard the stop event, resume the thread, and
# continue to wait for the thread to stop.... which it now never
# would.
#
# In order to try and expose this issue reliably, this test preloads a
# library that intercepts waitpid calls. All waitpid calls targeting
# pid -1 with the WNOHANG flag are rate limited so that only 1 per
# second can complete. Additional calls are forced to return 0
# indicating no event waiting. This is enough to trigger the bug
# during the attach phase.
# This test only works on Linux
if { ![isnative] || [is_remote host] || [use_gdb_stub]
|| ![istarget *-linux*] } {
continue
}
standard_testfile
set libfile slow-waitpid
set libsrc "${srcdir}/${subdir}/${libfile}.c"
set libobj [standard_output_file ${libfile}.so]
with_test_prefix "compile preload library" {
# Compile the preload library. We only get away with this as we
# limit this test to running when ISNATIVE is true.
if { [gdb_compile_shlib_pthreads \
$libsrc $libobj {debug}] != "" } then {
return -1
}
}
with_test_prefix "compile test executable" {
# Compile the test program
if { [gdb_compile_pthreads \
"${srcdir}/${subdir}/${srcfile}" "${binfile}" \
executable {debug}] != "" } {
return -1
}
}
# Spawn GDB with LIB preloaded with LD_PRELOAD.
proc gdb_spawn_with_ld_preload {lib} {
global env
save_vars { env(LD_PRELOAD) } {
if { ![info exists env(LD_PRELOAD) ]
|| $env(LD_PRELOAD) == "" } {
set env(LD_PRELOAD) "$lib"
} else {
append env(LD_PRELOAD) ":$lib"
}
gdb_start
}
}
# Run test program in the background.
set test_spawn_id [spawn_wait_for_attach $binfile]
set testpid [spawn_id_get_pid $test_spawn_id]
# Start GDB with preload library in place.
gdb_spawn_with_ld_preload $libobj
# Load binary, and attach to running program.
gdb_load ${binfile}
gdb_test "attach $testpid" "Attaching to program.*" "attach to target"
gdb_exit
# Kill of test program.
kill_wait_spawned_process $test_spawn_id