On Thu, 18 Jan 2024 16:19:38 GMT, Chris Plummer <cjplum...@openjdk.org> wrote:
>> I noticed that "clhsdb jstack" seemed to hang when I attached to process >> with a somewhat large heap. It had taken over 10 minutes when I finally >> decided to have a look at the SA process (using bin/jstack), which came up >> with the following in the stack: >> >> >> at >> sun.jvm.hotspot.oops.ObjectHeap.iterateObjectsOfKlass([jdk.hotspot.agent@23-internal](mailto:jdk.hotspot.agent@23-internal)/ObjectHeap.java:117) >> at >> sun.jvm.hotspot.runtime.ConcurrentLocksPrinter.fillLocks([jdk.hotspot.agent@23-internal](mailto:jdk.hotspot.agent@23-internal)/ConcurrentLocksPrinter.java:70) >> at >> sun.jvm.hotspot.runtime.ConcurrentLocksPrinter.<init>([jdk.hotspot.agent@23-internal](mailto:jdk.hotspot.agent@23-internal)/ConcurrentLocksPrinter.java:36) >> at >> sun.jvm.hotspot.tools.StackTrace.run([jdk.hotspot.agent@23-internal](mailto:jdk.hotspot.agent@23-internal)/StackTrace.java:71) >> >> >> This code is meant to print information about j.u.c. locks. It it works by >> searching the entire java heap for instances of AbstractOwnableSynchronizer. >> This is a very expensive operation because it means not only does SA need to >> read in the entire java heap, but it needs to create a Klass mirror for >> every heap object so it can determine its type. >> >> Our testing doesn't seem to run into this slowness problem because "clhsdb >> jstack" only iterates over the heap if AbstractOwnableSynchronizer has been >> loaded, which it won't be if no j.u.c. locks have been created. This seems >> to be the case wherever we use "clhsdb jstack" in testing. We do have some >> tests that likely result in j.u.c locks, but they all use "jhsdb jstack", >> which doesn't have this issue (it requires use of the --locks argument to >> enable printing of j.u.c locks). >> >> This issue was already called out in >> [JDK-8262098](https://bugs.openjdk.org/browse/JDK-8262098). For this CR I am >> proposing that "clhsdb jstack" not include j.u.c lock scanning by default, >> and add the -l argument to enable it. This will make it similar to >> bin/jstack, which also has a -l argument, and to "clhsdb jstack", which has >> a --locks argument (which maps internally to the Jstack.java -l argument). >> >> The same changes are also being made to "clhsdb pstack". >> >> Tested with all tier1, tier2, and tier5 svc tests. > > Chris Plummer has updated the pull request incrementally with two additional > commits since the last revision: > > - update clhsdb docs > - update clhsdb docs This seems to be a change that will impact the end user of these tools. Instead of disabling j.u.c.locks by default and proving an option to turn them on, you should provide an option to turn them off, and our tests can use that. What you proposed makes sense for an initial design of the tool, but not for a change after the fact. ------------- PR Review: https://git.openjdk.org/jdk/pull/17479#pullrequestreview-1831190402
