harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Leo Li" <liyilei1...@gmail.com>
Subject Re: [classlib][nio]FileChannel and direct buffer reallocation(was Re: [jira] Commented: (HARMONY-4081) [classlib][nio] FileChannel.write(ByteBuffer[]) sometimes works incorrectly)
Date Tue, 11 Sep 2007 12:54:40 GMT
On 9/11/07, Mikhail Markov <mikhail.a.markov@gmail.com> wrote:
> Leo,
>
> I've tried it on my laptop without Hyperthreading - still it crashes.
> And, btw, it did not crashed before r567561 commit - could that be a problem
> in that patch?

In that patch, I just explicitly invoke System.gc when memory is
tight. It will aggravate the problem, but I am not sure whether it is
the root cause.

It will take some time to investigate it.

Good luck!
>
> Thanks,
> Mikhail
>
> On 9/11/07, Leo Li <liyilei1979@gmail.com> wrote:
> >
> > On 9/10/07, Mikhail Markov <mikhail.a.markov@gmail.com> wrote:
> > > Hi, Leo,
> > >
> > > I've just checked only FileChannelImpl changes with the latest svn
> > snapshot
> > > and got VM crash (in IBM VME) in threadstart WIN API function. Could you
> > > repeat this?
> >
> > Hi, Mikhail
> >     I have met this problem before. It seems the native block
> > allocated for the direct byte buffer is released before we expected so
> > the WIN API will reference an invalid address although the direct byte
> > buffer should have been pinned in the patch.
> >    After some studying, I cannot find obvious problem in the native
> > block reallocation mechanism in the class library. Actually IBM VME
> > and DRLVM both encounter the failure.
> >    But what makes me puzzle most is the problem can occur on RI
> > albeit with a much lower frequency. (Not sure whether it can be
> > reproduce on every machine.)Seems it is a cross classlib and cross vm
> > problem.:)
> >    I have got some hint but I have no proof so I was hesitating to
> > tell it in public: I have once shutdown the hyper-threading option and
> > then everything is ok. I will try to find a multi-processor machine
> > with hyper-threading shutdown for test to determine whether it is
> > related to hyper-threading or parallel multi-threading(not the style
> > of time slice sharing).
> >   Could you please also try this on your server with hyper-threading
> > closed since the result is always different on each machine?
> >
> > Good luck!
> >
> > >
> > > Thanks,
> > > Mikhail
> > >
> > > On 8/22/07, Leo Li <liyilei1979@gmail.com> wrote:
> > > >
> > > > Hi, Mikhail:
> > > >     I have just focused on the problem about how to ensure direct byte
> > > > buffer to be release in time. And after I applied the patch at
> > r567561,
> > > > although the problem of releasing direct byte buffer seems resolved, I
> > > > found
> > > > one testcase in FileChannel failed.
> > > >     After some studying I found the problem is coincident with
> > > > HARMONY-4081, in that there is a bug in FileChannel.write() that there
> > is
> > > > no
> > > > holder for temporarily allocated direct buffers then the they might be
> > > > gc-ed
> > > > and the related memory resource reallocated. My patch, which  might
> > > > intrigue GC, aggravates this problem and leads to failure in normal
> > > > testsuite.
> > > >     So could you please first commit the part of FileChannel.write()
> > and
> > > > let current tests pass?
> > > >
> > > > Thanks.
> > > >
> > > > On 8/21/07, Mikhail Markov <mikhail.a.markov@gmail.com> wrote:
> > > > >
> > > > > Thanks for detailed comments!
> > > > > You are right about the memory auto-freeing, so my modifications
of
> > > > > AbstractMemorySpy are not correct.
> > > > > See my comments for MappedByteBuffer below inlined.
> > > > >
> > > > > Still the changes in FileChannelImpl alone do not work: I've just
> > > > > re-tried:
> > > > > the test still fails and the following messages starts printing:
> > > > > ...
> > > > > Memory Spy! Fixed attempt to free memory that was not allocated
> > > > > PlatformAddress[29968352]
> > > > > ...
> > > > > I've added debug stack-trace printing and found that these messages
> > are
> > > > > printed when tried to free DirectBuffers at the end of
> > > > > FileChannelImpl.write()
> > > > > method. It's strange at least, that we could not explicitly free
> > > > > DirectBuffer which we allocated.
> > > > > Seems like these buffers were freed in
> > AbstractMemorySpy.orphanedMemory
> > > > ()
> > > > > method.
> > > > > The comment for DirectByteBuffer.free() method says:
> > > > > "Explicitly free the memory used by this direct byte buffer. If the
> > > > memory
> > > > > has already been freed then this is a no-op.
> > > > > ...
> > > > > Note this is is possible that the memory is freed by code that
> > reaches
> > > > > into
> > > > > the address and explicitly frees it 'beneith' us -- this is bad
> > form."
> > > > > Does it mean that freeing in AbstractMemorySpy.orphanedMemory() is
> > "bad
> > > > > form"? :-)
> > > > > We should somehow "synchronize" the explicit memory freeing and
> > > > > auto-freeing
> > > > > in AbstractMemorySpy.
> > > > > Looking into the code, i could propose to add additional boolean
> > > > parameter
> > > > > to AbstractMemorySpy.free() method to indicate if warning message
> > should
> > > > > be
> > > > > printed or not but the main problem here is that reproducer still
> > fails
> > > > if
> > > > > modify just FileChannelImpl, which means that auto-freeing does not
> > work
> > > > > as
> > > > > expected. And I'm not quite understand why it's so.
> > > > > Do you have any ideas?
> > > > >
> > > > > Thanks,
> > > > > Mikhail
> > > > >
> > > > >
> > > > > On 8/17/07, Yang Paulex <paulex.yang@gmail.com> wrote:
> > > > > >
> > > > > > I'm forwarding this discussion to dev-list to make the discussion
> > > > > > easier:).
> > > > > > Please see my comments inline.
> > > > > >
> > > > > > 2007/8/16, Mikhail Markov (JIRA) <jira@apache.org>:
> > > > > > >
> > > > > > >
> > > > > > >     [
> > > > > > >
> > > > > >
> > > > >
> > > >
> > https://issues.apache.org/jira/browse/HARMONY-4081?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12520216
> > > > > > ]
> > > > > > >
> > > > > > > Mikhail Markov commented on HARMONY-4081:
> > > > > > > -----------------------------------------
> > > > > > >
> > > > > > > Paulex,
> > > > > > >
> > > > > > > Thanks for the patch review!
> > > > > > >
> > > > > > > In the beginning, i've created the patch withouth
> > > > > MappedPlatformAddress
> > > > > > > and AbstractMemorySpy modifications, but this lead to
> > exceptions.
> > > > > Seems
> > > > > > like
> > > > > > > after explicit temporary buffers freeing at the end of
> > > > > > > FileChannelImpl.write() method, another attempt to free
the same
> > > > > > resources
> > > > > > > is made in RuntimeMemorySpy.alloc() method.
> > > > > > >
> > > > > > > About MappedPlatformAddress modification: yes - on Linux
it's as
> > you
> > > > > > > described, but unfortunately on Windows UnmapViewOfFile
function
> > is
> > > > > > used,
> > > > > > > which does not physically free the memory.
> > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > I missed to mention windows implementation last time, but I
don't
> > > > catch
> > > > > > you
> > > > > > up here on the UnmapViewOfFile, because I cannot find the relevant
> > > > > > explanation in MSDN that this method needs further free() to
> > release
> > > > > > physical memory [1], and the sample code of MSDN doesn't add
any
> > > > further
> > > > > > memory free for this[2]. Even if UnmapViewOfFile doesn't free
the
> > > > > memory,
> > > > > > I
> > > > > > prefer we modify the unmap() implementation on Windows to add
> > memory
> > > > > free,
> > > > > > so that the the platform neutral behavior can be kept for portlib,
> > > > > > otherwise
> > > > > > on Linux we may put the situation in risk that free same memory
> > twice.
> > > > > how
> > > > > > do you think?
> > > > >
> > > > >
> > > > > I've just checked the info again and agree with you - no explicit
> > memory
> > > > > freeing is needed.
> > > > >
> > > > > But I did see some potential problems, although not sure because
> > MSDN is
> > > > > not
> > > > > > very clear here:-
> > > > > >
> > > > > > The CloseHandle() needs to be invoked after all file view is
> > unmapped,
> > > > > in
> > > > > > our case, we don't support multi-view for same file mapping
> > object, so
> > > > > > it's
> > > > > > OK to close the handle right after unmap. But for some unknown
> > > > reasons,
> > > > > > currently CloseHandle() is done in windows version's mmapImpl(Ln.
> > 151,
> > > > > > luni/src/main/native/luni/windows/OSMemoryWin32.c) right after
> > > > > > MapViewOfFile, I'm not sure if this is right action or not.
Some
> > > > > relevant
> > > > > > MSDN pages:
> > > > > >
> > > > > > "Unmapping a file view invalidates the pointer to the process's
> > > > virtual
> > > > > > address space. If any of the pages of the file view have changed
> > since
> > > > > the
> > > > > > view was mapped, the system writes the changed pages of the
file
> > to
> > > > disk
> > > > > > using caching. To commit all data to disk before unmapping the
> > file
> > > > > view,
> > > > > > use the *FlushViewOfFile*<
> > > > > > http://msdn2.microsoft.com/en-us/library/aa366563.aspx>function.
> > > > > >
> > > > > > When each process finishes using the file mapping object and
has
> > > > > unmapped
> > > > > > all views, it must close the file mapping object's handle and
the
> > file
> > > > > on
> > > > > > disk by calling
> > > > > > *CloseHandle*<
> > http://msdn2.microsoft.com/en-us/library/ms724211.aspx>.
> > > > > > These calls to *CloseHandle* succeed even when there are file
> > views
> > > > that
> > > > > > are
> > > > > > still open. However, leaving file views mapped causes memory
> > leaks."
> > > > > > In Harmony implementation, we actually call CloseHandle before
the
> > > > only
> > > > > > mapped file view is unmapped, but from the document above, I
> > cannot
> > > > say
> > > > > > it's
> > > > > > safe or not. I'll try to find some time to test later.
> > > > >
> > > > >
> > > > > I've read in [1] in Remarks:
> > > > > "Although an application may close the file handle used to create
a
> > file
> > > > > mapping object, the system holds the corresponding file open until
> > the
> > > > > last
> > > > > view of the file is unmapped: Files for which the last view has not
> > yet
> > > > > been
> > > > > unmapped are held open with no sharing restrictions."
> > > > >
> > > > > So, seems like immediate closing file handles after mapping looks
> > ok...
> > > > >
> > > > >
> > > > > [1] http://msdn2.microsoft.com/en-us/library/aa366882.aspx
> > > > > > [2] http://msdn2.microsoft.com/en-us/library/aa366548.aspx
> > > > > > [3] http://msdn2.microsoft.com/en-us/library/aa366532.aspx
> > > > > >
> > > > > > About AbstractMemorySpy modification: the modification is related
> > to
> > > > the
> > > > > > one
> > > > > > > in MappedPlatformAddress. Usually, the following construction
is
> > > > used
> > > > > > for
> > > > > > > resources explicit freeing:
> > > > > > >         if(memorySpy.free(this)){
> > > > > > >             osMemory.free(osaddr);
> > > > > > >         }
> > > > > > > i.e. after removing the address from memoryInUse, physical
> > freeing
> > > > > > happens
> > > > > > > - in this case. The only place where this was not so is
> > > > > > > MappedPlatformAddress (at least for Windows), so after
i added
> > > > > explicit
> > > > > > > memory freeing in MappedPlatformAddress, the address could
be
> > safely
> > > > > > removed
> > > > > > > from refToShadow.
> > > > > > > You're right - is this case the mechanizm of auto-freeing
is not
> > > > > > > necessary.
> > > > > > > I did not found places where free() method explicity used
> > > > > > > except  *PlatformAddress classes. Do you know any?
> > > > > > > If not then do we really need this mechanizm if all physical
> > freeing
> > > > > is
> > > > > > > done in *PlatformAddress classes?
> > > > > >
> > > > > >
> > > > > > PlatformAddress is used not only by MappedDirectBuffer but by
> > common
> > > > > > direct
> > > > > > buffer, too. We cannot ask applications running on Harmony to
> > > > explicitly
> > > > > > free all direct buffer, so the automatic reallocation  mechanism
> > is
> > > > > still
> > > > > > necessary.  If the number/size of direct buffer used by
> > > > > FileChannelImpl's
> > > > > > gather/scatter IO make you uncomfortable so that they are expected
> > to
> > > > be
> > > > > > released explicitly and quickly, I prefer we find some way to
deal
> > > > with
> > > > > > this
> > > > > > within FileChannelImpl rather than in a method of PlatformAddress
> > or
> > > > > > AbstractMemorySpy, which may be depended on by other classlib
and
> > in
> > > > > turn
> > > > > > by
> > > > > > applications. How do you think?
> > > > > >
> > > > > > --
> > > > > > Paulex Yang
> > > > > > China Software Development laboratory
> > > > > > IBM
> > > > > >
> > > > >
> > > >
> > > >
> > > >
> > > > --
> > > > Leo Li
> > > > China Software Development Lab, IBM
> > > >
> > >
> >
> >
> > --
> > Leo Li
> > China Software Development Lab, IBM
> >
>


-- 
Leo Li
China Software Development Lab, IBM

Mime
View raw message