Solaris specific bug in GC log rotation - extremely long pause

[Rainer Jung]

In case anyone is interested, I'll attach some reproduction code.
You can run it as a standalone Java program on Solaris. It is pretty 
small and simple. A README is included.

I hope the list accepts attachments.

Regards,

Rainer

Am 10.11.2015 um 16:45 schrieb Rainer Jung:
> There's a bug in the Solaris specific HotSpot code which exhibits itself
> as extremely long pauses directly after GC log rotation.
>
> The bug can be seen in 1.7.0 patchlevel 80 and above, 1.8.0 patchlevel
> 20 and above and 1.9.0. It was noticed using
> -XX:PrintGCApplicationStoppedTime.
>
> During GC log rotation the function os::print_memory_info() in
> src/os/solaris/vm/os_solaris.cpp is called, which itself calls
> check_addr0().
>
> Now check_addr0() reads the file /proc/self/pmap, but does this entry by
> entry. Productive Java processes often have 1000 and more map entries
> (e.g. each thread stack is one entry, often each jar etc.
>
> Reading one entry takes about 35-40ms in a zone on a T4 system. Reading
> e.g. 1000 entries takes 35-40 seconds during which the JVM is stopped.
> We experienced even stop times up until 2400 seconds in extreme cases.
>
> Note that this doesn't have to do with GC. It only happens directly
> after GC log file rotation.
>
> The bug is reading the file entry by entry instead of using pread() to
> read it in one go. Look for instance at the source code of the Solaris
> command "pmap" which uses pread() to read the file at once.
>
> It might also be questioned, whether this map file check must happen
> during a safepoint or could also be done outside, only synchronizing in
> case a hit was found and a warning actually needs to be written to the
> log file.
>
> The bug was shadowed by another bug until and including 1.7.0_76 and
> 1.8.0_11: the close() that closes the map file after reading the entries
> was erroneously put inside the read loop instead of after the read loop.
> Thus these versions only read the first entry, then closed the file and
> erred out when trying to read the next entry. So reading was quick.
>
> The responsible change was:
>
> Java 8:
> http://hg.openjdk.java.net/jdk8u/jdk8u/hotspot/diff/7d28f4e15b61/src/os/solaris/vm/os_solaris.cpp
>
>
> Java 7:
> http://hg.openjdk.java.net/jdk7u/jdk7u/hotspot/diff/e50eb3195734/src/os/solaris/vm/os_solaris.cpp
>
>
> Example code for reading using pread() (going back to OpenSolaris):
>
> https://github.com/joyent/illumos-joyent/blob/master/usr/src/cmd/ptools/pmap/pmap.c
>
>
> Look e.g. at function xmapping_iter() and there look for pread().
>
> See also the discussion thread starting at
> http://mail.openjdk.java.net/pipermail/hotspot-gc-use/2015-November/002351.html.
>
>
> The analysis was done using source code review and realtime checks with
> pstack and truss. It would be nice if you could instruct me how to get
> this fixed. It is a bit specific because it only happens on Solaris and
> when GC log rotation is active, but Solaris should be an important
> platform for Oracle and GC log rotation will be used more and more.
>
> I do *not* have a support contract.
>
> Regards,
>
> Rainer
-------------- next part --------------
A non-text attachment was scrubbed...
Name: GcRotationPause.tar.gz
Type: application/gzip
Size: 117137 bytes
Desc: not available
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20151110/9411d267/GcRotationPause.tar.gz>