BUG: withOwnerThread closes MappedMemorySegment
Ty Young
youngty1997 at gmail.com
Thu Sep 24 03:01:10 UTC 2020
On 9/23/20 4:31 AM, Maurizio Cimadamore wrote:
> A lot of methods in MemorySegment behave this way - this has always
> been the behavior of "withOwnerThread" even before we had shared
> segments.
withAccessModes didn't when I swapped withOwnerThread with it nor did
the previous restricted escape hatch method, which is why my code worked
before and thought it was a bug.
>
> There are two main issues with "just mutating the existining segments":
>
> 1) First and foremost, C2 likes constants. So, if the thread is
> constant on the segment we pay nothing for the ownership (or lack of
> ownership) check.
>
> 2) Second, you can't just "flip" a switch on a shared data structure,
> and expect the update to be seen by the rest of the world (in a
> multi-thread scenario). On the other hand, publishing a "new" object
> has more guarantees on who can do what with that new object (meaning
> it's physically not possible for other threads to start writing on the
> new segment before it has been returned by the API)
>
> Note that the original segment is killed (meaning its isAlive returns
> false) but the new segment still points at the same mappped memory
> region. So if you want a shared memory segment, you just have to do:
>
> MappedMemorySegment nativeFile =
> MemorySegment.mapFromPath(file.toPath(), 0, 4096,
> FileChannel.MapMode.READ_WRITE)
> .withOwnerThread(null);
Yeah, figured that'd work. I had to add an if-else to check if a segment
is already unbound, which got me past that error, although my projects
still aren't running because of a StackOverflow and
"java.lang.NoClassDefFoundError: Could not initialize class
java.lang.StackTraceElement$HashedModules" error related to
MethodHandles that are being added to struct layouts. Not 100% sure
what's going on there...
>
> E.g. use chaining to get to the segment you want to construct.
>
> As for the interaction with asSlice() - a slice is not a true
> standalone segment - is just another view of the existing segment with
> different bounds - in fact closing the slice also closes the parent
> segments. So, calling withOwnerThread will also kill all associated
> slices (or, calling withCleanupAction will attach the cleanup action
> to all the segments which share the same temporal bound).
>
> There is, currently, no way to create a segment, and then slice it so
> that you can N segments each owned by a different thread, if that's
> what you are looking for. We looked into that and it's way too messy
> (not only you have to worry about how to "split" the segment, but also
> about how you merge it back - and that's the hard part). So, if you
> are not ok with confinement, create a shared segment and work with it
> (e.g. like it was a ByteBuffer or a Java array). It is up to you then
> to make sure that multiple threads operate on the segment in a way
> that makes sense (either by assigning disjoint slices to different
> threads, e.g. using a spliterator, which the API supports), or by
> using some synchronization.
Figured I'd have to do manual synchronization.
>
> Maurizio
>
>
> On 23/09/2020 07:47, sundararajan.athijegannathan at oracle.com wrote:
>> Can you access source?
>>
>> https://github.com/openjdk/panama-foreign/blob/foreign-jextract/src/jdk.incubator.foreign/share/classes/jdk/incubator/foreign/MemorySegment.java
>>
>>
>> Yes, afaik it is implementation technical reason. (Maurizio will
>> clarify). That said, it should be possible to map different file
>> segments to *different* memory segments with different owners, right?
>> What's the advantage of mapping the whole & slice to own parts by
>> different threads?
>>
>> -Sundar
>>
>> On 23/09/20 11:02 am, Ty Young wrote:
>>> Javadoc isn't working for withOwnerThread or withCleanupAction for
>>> me on Netbeans:
>>>
>>>
>>> https://imgur.com/a/Q21pZAC
>>>
>>>
>>> Says it's downloading HTTP Javadoc for about 10 seconds then gives
>>> up. Javadoc of other older methods work just fine and I didn't think
>>> it'd close the segment as other withers(withAccessModes) do not.
>>>
>>>
>>> So I wasn't able to read it, my bad.
>>>
>>>
>>> Question: is this an absolute requirement because of technical
>>> restrains? Could this be removed so that, for example, it is
>>> possible to make an array segment unbound but have individual array
>>> index segments be bound to a specific thread?
>>>
>>>
>>> On 9/23/20 12:00 AM, sundararajan.athijegannathan at oracle.com wrote:
>>>> javadoc of MemorySegment.withOwnerThread starts as follows:
>>>>
>>>> " * Obtains a new memory segment backed by the same underlying
>>>> memory region as this segment,
>>>> * but with different owner thread. As a side-effect, this
>>>> segment will be marked as <em>not alive</em>,
>>>> * and subsequent operations on this segment will result in
>>>> runtime errors.
>>>> "
>>>>
>>>> So the behavior seen is as the specification.
>>>>
>>>> -Sundar
>>>>
>>>> On 23/09/20 7:36 am, Ty Young wrote:
>>>>> A bug seems to have been introduced wherein using withOwnerThread
>>>>> causes a MappedMemorySegment to close:
>>>>>
>>>>>
>>>>> File file = new File("./test");
>>>>>
>>>>> if(file.exists())
>>>>> file.delete();
>>>>>
>>>>> file.createNewFile();
>>>>>
>>>>> MappedMemorySegment nativeFile =
>>>>> MemorySegment.mapFromPath(file.toPath(), 0, 4096,
>>>>> FileChannel.MapMode.READ_WRITE);
>>>>>
>>>>> System.out.println(nativeFile.isAlive());
>>>>>
>>>>> MappedMemorySegment segment =
>>>>> (MappedMemorySegment)nativeFile.asSlice(4);
>>>>> segment = (MappedMemorySegment)nativeFile.asSlice(4);
>>>>> segment =
>>>>> (MappedMemorySegment)nativeFile.asSlice(4).withOwnerThread(null);
>>>>>
>>>>> System.out.println(nativeFile.isAlive());
>>>>> System.out.println(segment.isAlive());
>>>>>
>>>>>
>>>>> which prints true, false, and then true. The original
>>>>> MappedMemorySegment was never closed, so this is unexpected.
>>>>>
More information about the panama-dev
mailing list