Re: RFR (M): 8129419: heapDumper.cpp: assert(length_in_bytes > 0) failed: nothing to copy

[Andreas Eriksson]

Hi,

New webrev: http://cr.openjdk.java.net/~aeriksso/8129419/webrev.02/hotspot/
Write_internal handles writes less than len, and EINTR.

EINTR is taken care of by os::write, but to use it I had to remove an 
assert from the solaris os::write, which checked that a JavaThread was 
in_native. Since heap dumping is done from the vm_thread (not a 
JavaThread) the assert crashed the VM.

Regards,
Andreas

On 2016-01-13 13:39, Andreas Eriksson wrote:
Hi,

On 2015-12-29 21:27, Dmitry Samersoff wrote:
Andreas,

Great work. All but write_internal looks good for me (see below).


HprofReader.java:

Nit -  length -= skipped; should be after if skipped == 0

heapDumper.cpp:480

1. For windows you can do :
     assert(len < (size_t)UINT_MAX, ... );
     ::write( ..., (uint) (len & UINT_MAX));

2. You should check for n < len and if it is true,
deal with errno. n = 0 is legitimate case,
so assert is not necessary.

I only think n = 0 is valid if write is called with length 0, which 
write_internal will never do.
Otherwise, according to posix, n will be -1 if an error occurred, or 
greater than zero if some bytes were written. [1]
If some bytes but not all were written then the while(len > 0) loop 
will make sure we try to write the rest of the bytes.
It looks like windows write should never return 0 when len != 0 
either. [2]

I should however handle the -1 result with errno EINTR, working on a 
new webrev.

Thanks,
Andreas

[1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/write.html
If write() is interrupted by a signal before it writes any data, it 
shall return -1 with errno set to [EINTR].
If write() is interrupted by a signal after it successfully writes 
some data, it shall return the number of bytes written.

[2] https://msdn.microsoft.com/en-us/library/1570wh78(v=vs.80).aspx

-Dmitry

On 2015-12-28 17:02, Andreas Eriksson wrote:
Hi,

Here is the webrev for type changes.
Top-repo:
http://cr.openjdk.java.net/~aeriksso/8129419/webrev.01/jdk9-hs-rt/
Hotspot: 
http://cr.openjdk.java.net/~aeriksso/8129419/webrev.01/hotspot/

The windows_x64 native write function uses a length of type uint, not
size_t. I added an ifdef for that case and handled it, but better
suggestions are welcome.

Also found and fixed another problem in the hprof parser when skipping
over bytes.

I've not included the tests, they have OOM issues on several JPRT 
hosts,
and until I've figured out a way to fix that I wont be pushing them.

Thanks,
Andreas

On 2015-12-14 19:34, Andreas Eriksson wrote:
Hi Dmitry, thanks for looking at this.


On 2015-12-14 18:30, Dmitry Samersoff wrote:
Andreas,

Nice cleanup.

Some generic comments below.

1. It would be excellent if you can split webrev into two parts, one
part with types cleanup and other part with array truncation related
changes or ever push these changes separately as it address different
problems.

Type cleanup could be reviewed quickly, but review of array 
truncation
requires some time.

(We already have two different CRs: JDK-8129419 for type cleanup and
JDK-8144732 for array truncation)
Yes, that's a good point.


2.
it might be better to use size_t and jlong (or julong) across all 
file
and get rid of all other types with a few exclusion.
I'll see what I can do, and we can look at it closer once I've split
the type changes into a separate webrev.

3.
Each assert costs some time in nightly testing, so we probably don't
need as much asserts.
Alright.

4. write_internal:

    There are couple of cases where ::write writes less bytes than
requested and doesn't return -1.
    So we should check if ::write returns a value less that len 
and if it
happens deal with errno - either repeat entire write
attempt(EINTR/EAGAIN) or abort writing.
Actually, I meant to ask about that, but forgot:
I tried using os::write, which handles EINTR/EAGAIN if necessary
(depending on platform), but Solaris has an assert that fails:
http://hg.openjdk.java.net/jdk9/hs-rt/hotspot/file/5a42c1dde332/src/os/solaris/vm/os_solaris.cpp#l5692 


It checks that os::write is called by a JavaThread that is in_native,
which we are not, since we are in the VM_thread doing a safepoint_op.
Does anyone know if that assert is really needed? It's only done for
Solaris.

5. we should handle zero-length array in calculate_array_max_length -
it's a legitimate case
Not sure what you mean here, I believe it does handle zero-length 
arrays.
It should just fall through without taking any of the if-clauses.
(Or do you mean that I should add a return immediately at the top if
length is zero?)

6. max_juint is less than SegmentedHeapDumpThreshold so non-segmented
heapdump can't contain huge array and it's better to check for 
segmented
dump before any other checks.
Correct me if I'm wrong here, but SegmentedHeapDumpThreshold could in
theory be set so that we have a non-segmented heap dump while still
having huge arrays.
Not sure if this is worth taking into account, since it most likely
would lead to other problems as well, and the flag is develop only, so
it can't happen in product builds.
What do you think would be best to do here?

- Andreas

-Dmitry


On 2015-12-14 18:38, Andreas Eriksson wrote:
Hi,

Please review this fix for dumping of long arrays, and general 
cleanup
of types in heapDumper.cpp.

Problem:
At several places in heapDumper.cpp overflows could happen when 
dumping
long arrays.
Also the hprof format uses an u4 as a record length field, but 
arrays
can be longer than that (counted in bytes).

Fix:
Many types that were previously signed are changed to equivalent
unsigned types and/or to a larger type to prevent overflow.
The bulk of the change though is the addition of
calculate_array_max_length, which for a given array returns the 
number
of elements we can dump. That length is then used to truncate arrays
that are too long.
Whenever an array is truncated a warning is printed:
Java HotSpot(TM) 64-Bit Server VM warning: cannot dump array of type
object[] with length 1,073,741,823; truncating to length 536,870,908

Much of the rest of the change is moving functions needed by
calculate_array_max_length to the DumpWriter or DumperSupport 
class so
that they can be accessed.

Added a test that relies on the hprof parser, which also had a
couple of
overflow problems (top repo changes).
I've also run this change against a couple of other tests, but 
they are
problematic in JPRT because they are using large heaps and lots of
disk.

Bug:
8129419: heapDumper.cpp: assert(length_in_bytes > 0) failed: 
nothing to
copy
https://bugs.openjdk.java.net/browse/JDK-8129419

Also fixed in this change is the problems seen in these two bugs:
8133317: Integer overflow in heapDumper.cpp leads to corrupt HPROF
dumps
https://bugs.openjdk.java.net/browse/JDK-8133317

8144732: VM_HeapDumper hits assert with bad dump_len
https://bugs.openjdk.java.net/browse/JDK-8144732

Webrev:
Top repo:
http://cr.openjdk.java.net/~aeriksso/8129419/webrev.00/jdk9-hs-rt/
Hotspot:
http://cr.openjdk.java.net/~aeriksso/8129419/webrev.00/hotspot/

Regards,
Andreas