Hi Jini,
Thanks for investigating the size expansion issue.
I agree that the size increase is worth it. Even when not using SA, if
we open the core file inside GDB, we cannot read certain sections in the
CDS archive (such as the RO section and strings sections). That would
make debugging difficult. So I am in favor of this change.
For the JDK modules file, maybe we can unmap it in the signal handler,
before going ahead with the core dump? I think it's hardly needed for
debugging purposes. (Perhaps we can also do the same for the NIO mapped
files?)
A opt-flag as suggested by Kevin is a good idea.
Thanks
- Ioi
On 11/9/18 3:29 AM, Kevin Walls wrote:
Hi Jini,
Looks good to me. It might be a significant increase in size of
_some_ core files, but so many core files we see are much larger, in
gigabytes++ of course, so the CDS data size should not be such a
significant increase on (I think) most files.
The flexibiity of always having the CDS data there is very
significant. A core file should ideally be usable, without
additionally requiring the CDS archive from the machine. That
additional human round-trip upload request on every transmitted core
that needs investigating, seems like a less efficient route...).
Is there an opt-out? It's conditional on UseSharedSpaces but could
there be a flag to disable, in case we see crashes with gigabytes of
private mappings that we really don't want to retain (the user would
have to know to set a flag, to disable the new coredump filter ahead
of time).
Thanks!
Kevin
On 29/10/2018 06:02, Jini George wrote:
Thank you very much, Ioi, for looking into this, and the
clarification offline. My bad, I had missed the earlier mail from
you. :-( My responses below.
Yes, I had tested this on MacOS. The issue does not exist on MacOS
since the file backed private mmap()-ed regions get dumped into the
MacOS corefiles by default.
The corefile sizes on Linux do increase due to this change. And the
increase would also include any file mapped using NIO with
MapMode.PRIVATE. The typical corefile size increase with this change
would include the following components at a high level:
* Any NIO file mapping with MapMode.PRIVATE.
* Any file mmap()-ed by any native library with MAP_PRIVATE.
* The read only CDS regions (ro and od): Of the order of a few MB.
* The shared strings CDS region. (typically less than 1 MB).
* 2 MB per native shared library (regions with ---p permissions
mapped by the dynamic linker for better alignment and for keeping
libraries efficiently shareable).
* The JDK 'modules' file. (About 140 MB).
So, without including any NIO mapping, I typically see around 250-300
MB increase in the corefile sizes. I agree that the size increase
could be a cause for concern, but for FWIW, these privately mapped
files get dumped into the corefile for MacOS too. And the corefile
sizes for the same program on MacOS are way larger (of the order of a
few GB as against about 300 MB on Linux (without the change)).
The advantage of fixing this by modifying the coredump_filter v/s
doing it in SA (by reading in more sections of the shared archive
file) is that this would benefit other debuggers like gdb also. (And
reduces the dependence on having the shared archive file being
available at the time of debugging). If folks still think this is a
cause for concern, I could make modifications to fix this by reading
in the regions from the shared archive file in the SA code. I also
wonder if it is worth taking a relook at the mapping types of the
various CDS regions also.
Thank you,
Jini.
On 10/22/2018 10:27 AM, Ioi Lam wrote:
Hi Jini,
Did you see my earlier reply? I might have sent it out during the
mail server outage days :-(
http://openjdk.5641.n7.nabble.com/RFR-S-JDK-8200613-SA-jstack-throws-UnmappedAddressException-with-a-CDS-core-file-td352681.html#a353026
Here was my reply again:
Hi Jini,
The changes looks good to me.
Have you tested this on MacOS? CDS heap support is also enabled on
MacOS. See macros.hpp:
#if INCLUDE_CDS && INCLUDE_G1GC && defined(_LP64) &&
!defined(_WINDOWS)
#define INCLUDE_CDS_JAVA_HEAP 1
Also, besides CDS, do we know how often other files will be mmaped
with
MAP_PRIVATE? Will users get huge core files because CDS is enabled? In
JDK 12, CDS will be enabled by default (see JDK-8202951), so all users
will be affected by the following line:
if (UseSharedSpaces) {
set_coredump_filter(FILE_BACKED_PVT_BIT);
}
Maybe you can run an big app such as Eclipse, trigger a core dump, and
compare the size of the core file before/after this change?
Thanks
- Ioi
Thanks
- Ioi
On 10/21/18 8:58 PM, Jini George wrote:
Gentle reminder!
Thanks,
- Jini
On 10/9/2018 11:31 AM, Jini George wrote:
Hello!
[Including runtime-dev since the changes are in runtime code]
Requesting reviews for:
Webrev: http://cr.openjdk.java.net/~jgeorge/8200613/webrev.01/
BugID: https://bugs.openjdk.java.net/browse/JDK-8200613
Issue: jhsdb jstack would throw an UnmappedAddressException with a
core file generated from a CDS enabled java process. This is seen
only with Linux and with G1GC, while trying to read in data from
the shared strings region (the closed archive heap space). This
region (which is a file backed private memory region) is not
dumped into the corefile for Linux. This, being a heap region (and
therefore being a read-write region) is also not read in from the
classes.jsa file in SA since only the read only regions are read
in while processing the core file. (The expectation being that the
read write regions are in the core file).
Proposed solution: The proposed solution is to have the
coredump_filter value corresponding to the CDS process to include
bit 2 (file-backed private memory), so that the file-backed
private memory region also gets dumped into the corefile. The
proposed fix is in src/hotspot/os/linux/os_linux.cpp.
Thanks,
Jini.