Just to clarify ...
On 18/10/2017 10:28 PM, David Holmes wrote:
On 18/10/2017 8:26 PM, serguei.spit...@oracle.com wrote:
Hi David,
Thank you for jumping to this review and helping Yasumasa to sort it out!
I've just discovered that this issue was already on the table for
several months without a significant progress.
On 10/18/17 02:48, David Holmes wrote:
Hi Serguei
On 18/10/2017 7:25 PM, serguei.spit...@oracle.com wrote:
Hi Yasumasa,
Sorry for a quite late participation.
I looked at the previous webrevs and think that this one is much
better.
Some concern is if we need any kind of synchronization here, e.g. CAS.
But it depends on the PerfMemory class usage.
Should we make the static variables '_initialized' and '_destroyed'
volatile?
For good measure - yes.
Also, the '_initialized' is set to 1 with:
159 OrderAccess::release_store(&_initialized, 1);
Should we do the same to set the '_destroyed'?:
200 _destroyed = true;
There is a benign initialization race but we need the release_store
to ensure all the data fields can be read if _initialized is seen as
true. But what is missing is a load_acquire() in is_initialized() to
ensure we synchronize with that store!
Yes, I noticed that the load_acquire() is missed. :|
There is also a potential for a destruction race (if multiple aborts
happens concurrently in different threads) but that also seems
benign. In this case there is no data being set so the store to
_destroyed does not need to be a release_store.
I'm not convinced yet this is benign as the PerfMemory::destroy() has
this call:
197 delete_memory_region();
Yes though most of its work ends up being no-ops.
Now, I started thinking about the asserts that call the is_useable().
Should they be returns instead?
I think this is a somewhat confused chunk of code. It's only
fractionally thread-safe yet once in use could be in use concurrently
with an aborting thread that calls destroy(). I don't think there is any
simple fix for this. If we're in the process of crashing does it really
matter if we trigger a secondary crash due to this?
It doesn't matter if we do:
assert(is_usable(),...);
// continue
or
if (!is_usable()) return;
// continue
because as soon as we have checked is_usable() and abort happening in
another thread may have changed that by calling destroy.
This code is basically broken if we hit an abort path instead of a
normal VM shutdown.
David
-----
The problems with this code go way beyond what Yasumasa is trying to
address with the JSnap problem and I would not want to put it back on
him to try and come up with an overall solution.
Then the is_destroyed() would better to have the load_acquire().
You could add a load_acquire and do the store_release. It certainly
would not hurt, but I don't think it would actually benefit anything
either.
Cheers,
David
Just interested to know what do you think on this.
Thanks,
Serguei
Cheers,
David
Thanks,
Serguei
On 10/18/17 00:39, Yasumasa Suenaga wrote:
Hi David,
Thank you for your comment.
I uploaded new webrev:
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.07/
Serguei, please comment about this :-)
Yasumasa
2017-10-18 16:09 GMT+09:00 David Holmes<david.hol...@oracle.com>:
Hi Yasumasa,
On 18/10/2017 4:34 PM, Yasumasa Suenaga wrote:
Hi David,
I don't think we need the extra fields, just ensure the existing
ones
can't
be accessed (other than by the tools) after destroy is called.
I've added PerfMemory::is_useable() to check whether we can
access to
PerfMemory.
I think this webrev prevent to access to PerfMemory after
destroy() call.
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.06/
This:
90 void PerfMemory::initialize() {
91
92 if (_prologue != NULL)
93 // initialization already performed
94 return;
shouldn't check _prologue, but is_initialized().
213 assert(is_useable(), "called before initialization");
-> "called before init or after destroy"
Could add a similar assert in PerfMemory::mark_updated().
Let's see what Serguei thinks. :)
Thanks,
David
Thanks,
Yasumasa
2017-10-18 13:44 GMT+09:00 David Holmes<david.hol...@oracle.com>:
On 18/10/2017 2:27 PM, Yasumasa Suenaga wrote:
Hi David,
2017-10-18 12:55 GMT+09:00 David Holmes<david.hol...@oracle.com>:
On 18/10/2017 12:37 PM, Yasumasa Suenaga wrote:
Hi David,
With your changes you no longer null out _prologue so the
assertion
would
now not fail and we'd proceed to access the deleted memory
region!
On Linux, PerfMemory::delete_memory_region() does not call
munmap()
for PerfMemory.
Perhaps not but there are still other actions that happen and
the point
is
we should not be able to continue to use PerfMemory once it
has been
destroyed (even if the destruction is only logical).
I received same comment from Dmitry in the past, but we couldn't
decide how should we do.
http://mail.openjdk.java.net/pipermail/serviceability-dev/2016-May/019728.html
In that discussion, I uploaded another webrev which adds other
fields
for
JSnap.
Is it suitable?
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.02/
I don't think we need the extra fields, just ensure the existing
ones
can't
be accessed (other than by the tools) after destroy is called.
I'm unclear why you no longer clear all the fields set during
initialization?
PerfMemory.java in jdk.hotspot.agent needs these field values.
`jhsdb jsnap --core` is failed if they are cleared.
I'm not familiar with these tools. When do we produce a core
file after
calling PerfMemory::destroy ?
PerfMemory::destroy() is called before aborting.
Ah - right. I assume we need to close off the perfdata file
before we
abort.
Thanks,
David
-----------------------
#0 perfMemory_exit ()
at
/usr/src/debug/java-1.8.0-openjdk-1.8.0.144-7.b01.fc26.x86_64/openjdk/hotspot/src/share/vm/runtime/perfMemory.cpp:80
#1 0x00007f99b091c949 in os::shutdown ()
at
/usr/src/debug/java-1.8.0-openjdk-1.8.0.144-7.b01.fc26.x86_64/openjdk/hotspot/src/os/linux/vm/os_linux.cpp:1483
#2 0x00007f99b091c980 in os::abort (dump_core=<optimized out>)
at
/usr/src/debug/java-1.8.0-openjdk-1.8.0.144-7.b01.fc26.x86_64/openjdk/hotspot/src/os/linux/vm/os_linux.cpp:1503
#3 0x00007f99b0b689c3 in VMError::report_and_die (
this=this@entry=0x7ffcacf40b50)
at
/usr/src/debug/java-1.8.0-openjdk-1.8.0.144-7.b01.fc26.x86_64/openjdk/hotspot/src/share/vm/utilities/vmError.cpp:1060
#4 0x00007f99b0926f04 in JVM_handle_linux_signal
(sig=sig@entry=11,
info=info@entry=0x7ffcacf40df0,
ucVoid=ucVoid@entry=0x7ffcacf40cc0,
abort_if_unrecognized=abort_if_unrecognized@entry=1)
at
/usr/src/debug/java-1.8.0-openjdk-1.8.0.144-7.b01.fc26.x86_64/openjdk/hotspot/src/os_cpu/linux_x86/vm/os_linux_x86.cpp:541
-----------------------
Thanks,
Yasumasa
But it seems to me that there are various checks of
_prologue that should really be checking is_initialized()
and/or
is_destroyed() as a guard.
Should I change all assertions for _prologue?
Assertions and direct guards. Checking _prologue is a
placeholder for
the
real check.
Thanks,
David
Thanks,
Yasumasa
2017-10-18 10:53 GMT+09:00 David
Holmes<david.hol...@oracle.com>:
Hi Yasumasa,
By chance we ran into this bug which I analysed yesterday:
https://bugs.openjdk.java.net/browse/JDK-8189390
We hit the assertion:
# Internal Error
(/open/src/hotspot/share/runtime/perfMemory.cpp:216),
pid=17874, tid=17875
# assert(_prologue != __null) failed: called before
initialization
#
which is misleading because it can fail if called before
initialization,
or
after PerfMemory::destroy has been called.
With your changes you no longer null out _prologue so the
assertion
would
now not fail and we'd proceed to access the deleted memory
region!
I'm unclear why you no longer clear all the fields set during
initialization? But it seems to me that there are various
checks of
_prologue that should really be checking is_initialized()
and/or
is_destroyed() as a guard.
Thanks,
David
On 16/10/2017 11:25 PM, Yasumasa Suenaga wrote:
PING:
Could you review it?
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.05/
Thanks,
Yasumasa
On 2017/10/03 13:18, Yasumasa Suenaga wrote:
Hi all,
I added gtest unit test case for this change in new webrev:
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.05/
Could you review it?
Thanks,
Yasumasa
2017-09-27 0:01 GMT+09:00 Yasumasa
Suenaga<yasue...@gmail.com>:
Hi all,
I uploaded new webrev to be adapted to jdk10/hs:
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.04/
Thanks,
Yasumasa
On 2017/09/21 7:45, Yasumasa Suenaga wrote:
PING:
Have you checked this issue?
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.03/
Yasumasa
On 2017/07/01 23:43, Yasumasa Suenaga wrote:
PING:
Have you checked this issue?
Yasumasa
On 2017/06/13 14:10, Yasumasa Suenaga wrote:
Hi all,
I want to discuss about JDK-8151815: Could not parse
core image
with
JSnap.
In last year, I found JSnap cannot parse coredump and
I've sent
review
request for it as JDK-8151815. However it has not been
reviewed
yet
[1].
We've discussed about safety implementation, but we
could not
get
consensus.
IMHO all SA tools should be handled java processes and
core
images,
and PerfCounter value is useful. So I fix this issue.
I uploaded new webrev for this issue. I think this
patch is
safety
because new flag PerfMemory::_destroyed guards double
free, and
all
members in PerfMemory is accessible (they are not
munmap'ed)
http://cr.openjdk.java.net/~ysuenaga/JDK-8151815/webrev.03/
Can you cooperate?
Thanks,
Yasumasa
[1]
http://mail.openjdk.java.net/pipermail/serviceability-dev/2016-April/019480.html
