On 6/15/20 9:28 PM, David Holmes wrote:
On 16/06/2020 10:57 am, Daniel D. Daugherty wrote:
On 6/15/20 7:19 PM, David Holmes wrote:
On 16/06/2020 8:40 am, Daniel D. Daugherty wrote:
On 6/15/20 6:14 PM, David Holmes wrote:
Hi Dan,
On 15/06/2020 11:38 pm, Daniel D. Daugherty wrote:
On 6/15/20 3:26 AM, David Holmes wrote:
On 15/06/2020 4:02 pm, Yasumasa Suenaga wrote:
Hi David,
On 2020/06/15 14:15, David Holmes wrote:
Hi Yasumasa,
On 15/06/2020 2:49 pm, Yasumasa Suenaga wrote:
Hi all,
I wonder why JvmtiEnvBase::get_object_monitor_usage()
(implementation of GetObjectMonitorUsage()) does not perform
at safepoint.
GetObjectMonitorUsage will use a safepoint if the target is
not suspended:
jvmtiError
JvmtiEnv::GetObjectMonitorUsage(jobject object,
jvmtiMonitorUsage* info_ptr) {
JavaThread* calling_thread = JavaThread::current();
jvmtiError err = get_object_monitor_usage(calling_thread,
object, info_ptr);
if (err == JVMTI_ERROR_THREAD_NOT_SUSPENDED) {
// Some of the critical threads were not suspended. go to
a safepoint and try again
VM_GetObjectMonitorUsage op(this, calling_thread, object,
info_ptr);
VMThread::execute(&op);
err = op.result();
}
return err;
} /* end GetObject */
I saw this code, so I guess there are some cases when
JVMTI_ERROR_THREAD_NOT_SUSPENDED is not returned from
get_object_monitor_usage().
Monitor owner would be acquired from monitor object at first
[1], but it would perform concurrently.
If owner thread is not suspended, the owner might be changed
to others in subsequent code.
For example, the owner might release the monitor before [2].
The expectation is that when we find an owner thread it is
either suspended or not. If it is suspended then it cannot
release the monitor. If it is not suspended we detect that and
redo the whole query at a safepoint.
I think the owner thread might resume unfortunately after
suspending check.
Yes you are right. I was thinking resuming also required a
safepoint but it only requires the Threads_lock. So yes the code
is wrong.
Which code is wrong?
Yes, a rogue resume can happen when the GetObjectMonitorUsage()
caller
has started the process of gathering the information while not at a
safepoint. Thus the information returned by GetObjectMonitorUsage()
might be stale, but that's a bug in the agent code.
The code tries to make sure that it either collects data about a
monitor owned by a thread that is suspended, or else it collects
that data at a safepoint. But the owning thread can be resumed
just after the code determined it was suspended. The monitor can
then be released and the information gathered not only stale but
potentially completely wrong as it could now be owned by a
different thread and will report that thread's entry count.
If the agent is not using SuspendThread(), then as soon as
GetObjectMonitorUsage() returns to the caller the information
can be stale. In fact as soon as the implementation returns
from the safepoint that gathered the info, the target thread
could have moved on.
That isn't the issue. That the info is stale is fine. But the
expectation is that the information was actually an accurate
snapshot of the state of the monitor at some point in time. The
current code does not ensure that.
Please explain. I clearly don't understand why you think the info
returned isn't "an accurate snapshot of the state of the monitor
at some point in time".
Because it may not be a "snapshot" at all. There is no atomicity**.
The reported owner thread may not own it any longer when the entry
count is read, so straight away you may have the wrong entry count
information. The set of threads trying to acquire the monitor, or wait
on the monitor can change in unexpected ways. It would be possible for
instance to report the same thread as being the owner, being blocked
trying to enter the monitor, and being in the wait-set of the monitor
- apparently all at the same time!
** even if the owner is suspended we don't have complete atomicity
because threads can join the set of threads trying to enter the
monitor (unless they are all suspended).
Consider the case when the monitor's owner is _not_ suspended:
- GetObjectMonitorUsage() uses a safepoint to gather the info about
the object's monitor. Since we're at a safepoint, the info that
we are gathering cannot change until we return from the safepoint.
It is a snapshot and a valid one at that.
Consider the case when the monitor's owner is suspended:
- GetObjectMonitorUsage() will gather info about the object's
monitor while _not_ at a safepoint. Assuming that no other
thread is suspended, then entry_count can change because
another thread can block on entry while we are gathering
info. waiter_count and waiters can change if a thread was
in a timed wait that has timed out and now that thread is
blocked on re-entry. I don't think that notify_waiter_count
and notify_waiters can change.
So in this case, the owner info and notify info is stable,
but the entry_count and waiter info is not stable.
Consider the case when the monitor is not owned:
- GetObjectMonitorUsage() will start to gather info about the
object's monitor while _not_ at a safepoint. If it finds a
thread on the entry queue that is not suspended, then it will
bail out and redo the info gather at a safepoint. I just
noticed that it doesn't check for suspension for the threads
on the waiters list so a timed Object.wait() call can cause
some confusion here.
So in this case, the owner info is not stable if a thread
comes out of a timed wait and reenters the monitor. This
case is no different than if a "barger" thread comes in
after the NULL owner field is observed and enters the
monitor. We'll return that there is no owner, a list of
suspended pending entry thread and a list of waiting
threads. The reality is that the object's monitor is
owned by the "barger" that completely bypassed the entry
queue by virtue of seeing the NULL owner field at exactly
the right time.
So the owner field is only stable when we have an owner. If
that owner is not suspended, then the other fields are also
stable because we gathered the info at a safepoint. If the
owner is suspended, then the owner and notify info is stable,
but the entry_count and waiter info is not stable.
If we have a NULL owner field, then the info is only stable
if you have a non-suspended thread on the entry list. Ouch!
That's deterministic, but not without some work.
Okay so only when we gather the info at a safepoint is all
of it a valid and stable snapshot. Unfortunately, we only
do that at a safepoint when the owner thread is not suspended
or if owner == NULL and one of the entry threads is not
suspended. If either of those conditions is not true, then
the different pieces of info is unstable to varying degrees.
As for this claim:
It would be possible for instance to report the same thread
as being the owner, being blocked trying to enter the monitor,
and being in the wait-set of the monitor - apparently all at
the same time!
I can't figure out a way to make that scenario work. If the
thread is seen as the owner and is not suspended, then we
gather info at a safepoint. If it is suspended, then it can't
then be seen as on the entry queue or on the wait queue since
it is suspended. If it is seen on the entry queue and is not
suspended, then we gather info at a safepoint. If it is
suspended on the entry queue, then it can't be seen on the
wait queue.
So the info instability of this API is bad, but it's not
quite that bad. :-) (That is a small mercy.)
Handshaking is not going to make this situation any better
for GetObjectMonitorUsage(). If the monitor is owned and we
handshake with the owner, the stability or instability of
the other fields remains the same as when SuspendThread is
used. Handshaking with all threads won't make the data as
stable as when at a safepoint because individual threads
can resume execution after doing their handshake so there
will still be field instability.
Short version: GetObjectMonitorUsage() should only gather
data at a safepoint. Yes, I've changed my mind.
Dan
David
-----
The only way to make sure you don't have stale information is
to use SuspendThread(), but it's not required. Perhaps the doc
should have more clear about the possibility of returning stale
info. That's a question for Robert F.
GetObjectMonitorUsage says nothing about thread's being suspended
so I can't see how this could be construed as an agent bug.
In your scenario above, you mention that the target thread was
suspended, GetObjectMonitorUsage() was called while the target
was suspended, and then the target thread was resumed after
GetObjectMonitorUsage() checked for suspension, but before
GetObjectMonitorUsage() was able to gather the info.
All three of those calls: SuspendThread(), GetObjectMonitorUsage()
and ResumeThread() are made by the agent and the agent should not
resume the target thread while also calling GetObjectMonitorUsage().
The calls were allowed to be made out of order so agent bug.
Perhaps. I was thinking more generally about an independent resume,
but you're right that doesn't really make a lot of sense. But when
the spec says nothing about suspension ...
And it is intentional that suspension is not required. JVM/DI and JVM/PI
used to require suspension for these kinds of get-the-info APIs. JVM/TI
intentionally was designed to not require suspension.
As I've said before, we could add a note about the data being
potentially
stale unless SuspendThread is used. I think of it like stat(2). You can
fetch the file's info, but there's no guarantee that the info is current
by the time you process what you got back. Is it too much motherhood to
state that the data might be stale? I could go either way...
Using a handshake on the owner thread will allow this to be fixed
in the future without forcing/using any safepoints.
I have to think about that which is why I'm avoiding talking about
handshakes in this thread.
Effectively the handshake can "suspend" the thread whilst the
monitor is queried. In effect the operation would create a
per-thread safepoint.
I "know" that, but I still need time to think about it and probably
see the code to see if there are holes...
Semantically it is no different to the code actually suspending the
owner thread, but it can't actually do that because suspends/resume
don't nest.
Yeah... we used have a suspend count back when we tracked internal and
external suspends separately. That was a nightmare...
Dan
Cheers,
David
Dan
Cheers,
David
Dan
JavaThread::is_ext_suspend_completed() is used to check thread
state, it returns `true` when the thread is sleeping [3], or
when it performs in native [4].
Sure but if the thread is actually suspended it can't continue
execution in the VM or in Java code.
This appears to be an optimisation for the assumed common case
where threads are first suspended and then the monitors are
queried.
I agree with this, but I could find out it from JVMTI spec - it
just says "Get information about the object's monitor."
Yes it was just an implementation optimisation, nothing to do
with the spec.
GetObjectMonitorUsage() might return incorrect information in
some case.
It starts with finding owner thread, but the owner might be
just before wakeup.
So I think it is more safe if GetObjectMonitorUsage() is called
at safepoint in any case.
Except we're moving away from safepoints to using Handshakes, so
this particular operation will require that the apparent owner
is Handshake-safe (by entering a handshake with it) before
querying the monitor. This would still be preferable I think to
always using a safepoint for the entire operation.
Cheers,
David
-----
Thanks,
Yasumasa
[3]
http://hg.openjdk.java.net/jdk/jdk/file/76a17c8143d8/src/hotspot/share/runtime/thread.cpp#l671
[4]
http://hg.openjdk.java.net/jdk/jdk/file/76a17c8143d8/src/hotspot/share/runtime/thread.cpp#l684
However there is still a potential bug as the thread reported
as the owner may not be suspended at the time we first see it,
and may release the monitor, but then it may get suspended
before we call:
owning_thread =
Threads::owning_thread_from_monitor_owner(tlh.list(), owner);
and so we think it is still the monitor owner and proceed to
query the monitor information in a racy way. This can't happen
when suspension itself requires a safepoint as the current
thread won't go to that safepoint during this code. However,
if suspension is implemented via a direct handshake with the
target thread then we have a problem.
David
-----
Thanks,
Yasumasa
[1]
http://hg.openjdk.java.net/jdk/jdk/file/76a17c8143d8/src/hotspot/share/prims/jvmtiEnvBase.cpp#l973
[2]
http://hg.openjdk.java.net/jdk/jdk/file/76a17c8143d8/src/hotspot/share/prims/jvmtiEnvBase.cpp#l996
