On Mon, 22 Jan 2024 16:37:14 GMT, Aleksey Shipilev <sh...@openjdk.org> wrote:
>> Currently we don't record dependencies on redefined methods (i.e. >> `evol_method` dependencies) in JIT compiled methods if none of the >> `can_redefine_classes`, `can_retransform_classes` or >> `can_generate_breakpoint_events` JVMTI capabalities is set. This means that >> if a JVMTI agent which requests one of these capabilities is dynamically >> attached, all the methods which have been JIT compiled until that point, >> will be marked for deoptimization and flushed from the code cache. For >> large, warmed-up applications this mean deoptimization and instant >> recompilation of thousands if not then-thousands of methods, which can lead >> to dramatic performance/latency drop-downs for several minutes. >> >> One could argue that dynamic agent attach is now deprecated anyway (see [JEP >> 451: Prepare to Disallow the Dynamic Loading of >> Agents](https://openjdk.org/jeps/451)) and this problem could be solved by >> making the recording of `evol_method` dependencies dependent on the new >> `-XX:+EnableDynamicAgentLoading` flag isntead of the concrete JVMTI >> capabilities (because the presence of the flag indicates that an agent will >> be loaded eventually). >> >> But there a single, however important exception to this rule and that's JFR. >> JFR is advertised as low overhead profiler which can be enabled in >> production at any time. However, when JFR is started dynamically (e.g. >> through JCMD or JMX) it will silently load a HotSpot internl JVMTI agent >> which requests the `can_retransform_classes` and retransforms some classes. >> This will inevitably trigger the deoptimization of all compiled methods as >> described above. >> >> I'd therefor like to propose to *always* and unconditionally record >> `evol_method` dependencies in JIT compiled code by exporting the relevant >> properties right at startup in `init_globals()`: >> ```c++ >> jint init_globals() { >> management_init(); >> JvmtiExport::initialize_oop_storage(); >> +#if INCLUDE_JVMTI >> + JvmtiExport::set_can_hotswap_or_post_breakpoint(true); >> + JvmtiExport::set_all_dependencies_are_recorded(true); >> +#endif >> >> >> My measurements indicate that the overhead of doing so is minimal (around 1% >> increase of nmethod size) and justifies the benefit. E.g. a Spring Petclinic >> application started with `-Xbatch -XX:+UnlockDiagnosticVMOptions >> -XX:+LogCompilation` compiles about ~11500 methods (~9000 with C1 and ~2500 >> with C2) resulting in an aggregated nmethod size of around ~40bm. >> Additionally recording `evol_method` dependencies only increases this size >> be about 400kb.... > >> Instead of unconditionally recording evol_method dependencies we could guard >> the recording by a new flag. But this would only make sense if that flag >> would be on by default and I don't know if such a flag is justified just for >> the rare (or non-existent?) cases where somebody wants to disable the >> recording of the dependencies. > > I think introducing a diagnostic flag is sensible here. If we figure out much > later that this solution comes with some other (worse) problems, the > diagnostic flag gives us the options to: a) clearly point at this addition as > the culprit; b) have the easily deployable solution to restore the original > behavior. > > For the change itself, we need to amend the comment near > `VM_RedefineClasses::flush_dependent_code` definition that talks about this > peculiar behavior, which now changes. Actually, maybe even the implementation > of `flush_dependent_code` should now trust (and assert) that all dependencies > are now recorded? @shipilev I filed https://bugs.openjdk.org/browse/JDK-8324318 which is related to this change. ------------- PR Comment: https://git.openjdk.org/jdk/pull/17509#issuecomment-1904402169
