src/hotspot/share/compiler/compileBroker.cpp

*** 2742,2751 ****
--- 2742,2752 ----
  //       That's a tradeoff which keeps together important blocks of output.
  //       At the same time, continuous tty_lock hold time is kept in check,
  //       preventing concurrently printing threads from stalling a long time.
  void CompileBroker::print_heapinfo(outputStream* out, const char* function, size_t granularity) {
    TimeStamp ts_total;
+   TimeStamp ts_global;
    TimeStamp ts;
  
    bool allFun = !strcmp(function, "all");
    bool aggregate = !strcmp(function, "aggregate") || !strcmp(function, "analyze") || allFun;
    bool usedSpace = !strcmp(function, "UsedSpace") || allFun;
*** 2771,2818 ****
    if (aggregate) {
      print_info(out);
    }
  
    // We hold the CodeHeapStateAnalytics_lock all the time, from here until we leave this function.
!   // That helps us getting a consistent view on the CodeHeap, at least for the "all" function.
    // When we request individual parts of the analysis via the jcmd interface, it is possible
    // that in between another thread (another jcmd user or the vm running into CodeCache OOM)
!   // updated the aggregated data. That's a tolerable tradeoff because we can't hold a lock
!   // across user interaction.
    ts.update(); // record starting point
!   MutexLockerEx mu1(CodeHeapStateAnalytics_lock, Mutex::_no_safepoint_check_flag);
!   out->cr();
!   out->print_cr("__ CodeHeapStateAnalytics lock wait took %10.3f seconds _________", ts.seconds());
!   out->cr();
  
    if (aggregate) {
-     // It is sufficient to hold the CodeCache_lock only for the aggregate step.
-     // All other functions operate on aggregated data - except MethodNames, but that should be safe.
-     // The separate CodeHeapStateAnalytics_lock protects the printing functions against
-     // concurrent aggregate steps. Acquire this lock before acquiring the CodeCache_lock.
-     // CodeHeapStateAnalytics_lock could be held by a concurrent thread for a long time,
-     // leading to an unnecessarily long hold time of the CodeCache_lock.
      ts.update(); // record starting point
!     MutexLockerEx mu2(CodeCache_lock, Mutex::_no_safepoint_check_flag);
!     out->cr();
!     out->print_cr("__ CodeCache lock wait took %10.3f seconds _________", ts.seconds());
!     out->cr();
  
      ts.update(); // record starting point
      CodeCache::aggregate(out, granularity);
!     out->cr();
!     out->print_cr("__ CodeCache lock hold took %10.3f seconds _________", ts.seconds());
!     out->cr();
    }
  
    if (usedSpace) CodeCache::print_usedSpace(out);
    if (freeSpace) CodeCache::print_freeSpace(out);
    if (methodCount) CodeCache::print_count(out);
    if (methodSpace) CodeCache::print_space(out);
    if (methodAge) CodeCache::print_age(out);
!   if (methodNames) CodeCache::print_names(out);
    if (discard) CodeCache::discard(out);
  
!   out->cr();
!   out->print_cr("__ CodeHeapStateAnalytics total duration %10.3f seconds _________", ts_total.seconds());
!   out->cr();
  }
--- 2772,2853 ----
    if (aggregate) {
      print_info(out);
    }
  
    // We hold the CodeHeapStateAnalytics_lock all the time, from here until we leave this function.
!   // That prevents other threads from destroying (making inconsistent) our view on the CodeHeap.
    // When we request individual parts of the analysis via the jcmd interface, it is possible
    // that in between another thread (another jcmd user or the vm running into CodeCache OOM)
!   // updated the aggregated data. We will then see a modified, but again consistent, view
!   // on the CodeHeap. That's a tolerable tradeoff we have to accept because we can't hold
!   // a lock across user interaction.
! 
!   // We should definitely acquire this lock before acquiring Compile_lock and CodeCache_lock.
!   // CodeHeapStateAnalytics_lock may be held by a concurrent thread for a long time,
!   // leading to an unnecessarily long hold time of the other locks we acquired before.
    ts.update(); // record starting point
!   MutexLockerEx mu0(CodeHeapStateAnalytics_lock);
!   out->print_cr("\n__ CodeHeapStateAnalytics lock wait took %10.3f seconds _________\n", ts.seconds());
! 
!   // Holding the CodeCache_lock protects from concurrent alterations of the CodeCache.
!   // Unfortunately, such protection is not sufficient:
!   // When a new nmethod is created via ciEnv::register_method(), the
!   // Compile_lock is taken first. After some initializations,
!   // nmethod::new_nmethod() takes over, grabbing the CodeCache_lock
!   // immediately (after finalizing the oop references). To lock out concurrent
!   // modifiers, we have to grab both locks as well in the described sequence.
!   //
!   // If we serve an "allFun" call, it is beneficial to hold CodeCache_lock and Compile_lock
!   // for the entire duration of aggregation and printing. That makes sure we see
!   // a consistent picture and do not run into issues caused by concurrent alterations.
!   bool should_take_Compile_lock   = !SafepointSynchronize::is_at_safepoint() &&
!                                     !Compile_lock->owned_by_self();
!   bool should_take_CodeCache_lock = !SafepointSynchronize::is_at_safepoint() &&
!                                     !CodeCache_lock->owned_by_self();
!   Mutex*   global_lock_1   = allFun ? (should_take_Compile_lock   ? Compile_lock   : NULL) : NULL;
!   Monitor* global_lock_2   = allFun ? (should_take_CodeCache_lock ? CodeCache_lock : NULL) : NULL;
!   Mutex*   function_lock_1 = allFun ? NULL : (should_take_Compile_lock   ? Compile_lock    : NULL);
!   Monitor* function_lock_2 = allFun ? NULL : (should_take_CodeCache_lock ? CodeCache_lock  : NULL);
!   ts_global.update(); // record starting point
!   MutexLockerEx mu1(global_lock_1, Mutex::_no_safepoint_check_flag);
!   MutexLockerEx mu2(global_lock_2, Mutex::_no_safepoint_check_flag);
!   if ((global_lock_1 != NULL) || (global_lock_2 != NULL)) {
!     out->print_cr("\n__ Compile & CodeCache (global) lock wait took %10.3f seconds _________\n", ts_global.seconds());
!     ts_global.update(); // record starting point
!   }
  
    if (aggregate) {
      ts.update(); // record starting point
!     MutexLockerEx mu11(function_lock_1, Mutex::_no_safepoint_check_flag);
!     MutexLockerEx mu22(function_lock_2, Mutex::_no_safepoint_check_flag);
!     if ((function_lock_1 != NULL) || (function_lock_2 != NULL)) {
!       out->print_cr("\n__ Compile & CodeCache (function) lock wait took %10.3f seconds _________\n", ts.seconds());
!     }
  
      ts.update(); // record starting point
      CodeCache::aggregate(out, granularity);
!     if ((function_lock_1 != NULL) || (function_lock_2 != NULL)) {
!       out->print_cr("\n__ Compile & CodeCache (function) lock hold took %10.3f seconds _________\n", ts.seconds());
!     }
    }
  
    if (usedSpace) CodeCache::print_usedSpace(out);
    if (freeSpace) CodeCache::print_freeSpace(out);
    if (methodCount) CodeCache::print_count(out);
    if (methodSpace) CodeCache::print_space(out);
    if (methodAge) CodeCache::print_age(out);
!   if (methodNames) {
!     if (allFun) {
!       // print_names() can only be used safely if the locks have been continuously held
!       // since aggregation begin. That is true only for function "all".
!       CodeCache::print_names(out);
!     } else {
!       out->print_cr("\nCodeHeapStateAnalytics: Function 'MethodNames' is only available as part of function 'all'");
!     }
!   }
    if (discard) CodeCache::discard(out);
  
!   if ((global_lock_1 != NULL) || (global_lock_2 != NULL)) {
!     out->print_cr("\n__ Compile & CodeCache (global) lock hold took %10.3f seconds _________\n", ts_global.seconds());
!   }
!   out->print_cr("\n__ CodeHeapStateAnalytics total duration %10.3f seconds _________\n", ts_total.seconds());
  }