8310015: ZGC: Unbounded asynchronous unmapping can lead to running out of address space

Reviewed-by: stefank, aboldtch

Author: fisk

src/hotspot/share/gc/x/xUnmapper.cpp

@@ -23,6 +23,7 @@
 
 #include "precompiled.hpp"
 #include "gc/shared/gc_globals.hpp"
+#include "gc/shared/gcLogPrecious.hpp"
 #include "gc/x/xList.inline.hpp"
 #include "gc/x/xLock.inline.hpp"
 #include "gc/x/xPage.inline.hpp"
@@ -35,6 +36,8 @@ XUnmapper::XUnmapper(XPageAllocator* page_allocator) :
     _page_allocator(page_allocator),
     _lock(),
     _queue(),
+    _enqueued_bytes(0),
+    _warned_sync_unmapping(false),
     _stop(false) {
   set_name("XUnmapper");
   create_and_start();
@@ -50,13 +53,50 @@ XPage* XUnmapper::dequeue() {
 
     XPage* const page = _queue.remove_first();
     if (page != nullptr) {
+      _enqueued_bytes -= page->size();
       return page;
     }
 
     _lock.wait();
   }
 }
 
+bool XUnmapper::try_enqueue(XPage* page) {
+  if (ZVerifyViews) {
+    // Asynchronous unmap and destroy is not supported with ZVerifyViews
+    return false;
+  }
+
+  // Enqueue for asynchronous unmap and destroy
+  XLocker<XConditionLock> locker(&_lock);
+  if (is_saturated()) {
+    // The unmapper thread is lagging behind and is unable to unmap memory fast enough
+    if (!_warned_sync_unmapping) {
+      _warned_sync_unmapping = true;
+      log_warning_p(gc)("WARNING: Encountered synchronous unmapping because asynchronous unmapping could not keep up");
+    }
+    log_debug(gc, unmap)("Synchronous unmapping " SIZE_FORMAT "M page", page->size() / M);
+    return false;
+  }
+
+  log_trace(gc, unmap)("Asynchronous unmapping " SIZE_FORMAT "M page (" SIZE_FORMAT "M / " SIZE_FORMAT "M enqueued)",
+                       page->size() / M, _enqueued_bytes / M, queue_capacity() / M);
+
+  _queue.insert_last(page);
+  _enqueued_bytes += page->size();
+  _lock.notify_all();
+
+  return true;
+}
+
+size_t XUnmapper::queue_capacity() const {
+  return align_up<size_t>(_page_allocator->max_capacity() * ZAsyncUnmappingLimit / 100.0, XGranuleSize);
+}
+
+bool XUnmapper::is_saturated() const {
+  return _enqueued_bytes >= queue_capacity();
+}
+
 void XUnmapper::do_unmap_and_destroy_page(XPage* page) const {
   EventZUnmap event;
   const size_t unmapped = page->size();
@@ -70,15 +110,9 @@ void XUnmapper::do_unmap_and_destroy_page(XPage* page) const {
 }
 
 void XUnmapper::unmap_and_destroy_page(XPage* page) {
-  // Asynchronous unmap and destroy is not supported with ZVerifyViews
-  if (ZVerifyViews) {
-    // Immediately unmap and destroy
+  if (!try_enqueue(page)) {
+    // Synchronously unmap and destroy
     do_unmap_and_destroy_page(page);
-  } else {
-    // Enqueue for asynchronous unmap and destroy
-    XLocker<XConditionLock> locker(&_lock);
-    _queue.insert_last(page);
-    _lock.notify_all();
   }
 }
 

src/hotspot/share/gc/x/xUnmapper.hpp

@@ -36,9 +36,14 @@ class XUnmapper : public ConcurrentGCThread {
   XPageAllocator* const _page_allocator;
   XConditionLock        _lock;
   XList<XPage>          _queue;
+  size_t                _enqueued_bytes;
+  bool                  _warned_sync_unmapping;
   bool                  _stop;
 
   XPage* dequeue();
+  bool try_enqueue(XPage* page);
+  size_t queue_capacity() const;
+  bool is_saturated() const;
   void do_unmap_and_destroy_page(XPage* page) const;
 
 protected:

src/hotspot/share/gc/z/shared/z_shared_globals.hpp

@@ -61,6 +61,13 @@
           "Uncommit memory if it has been unused for the specified "        \
           "amount of time (in seconds)")                                    \
                                                                             \
+  product(double, ZAsyncUnmappingLimit, 100.0, DIAGNOSTIC,                  \
+          "Specify the max amount (percentage of max heap size) of async "  \
+          "unmapping that can be in-flight before unmapping requests are "  \
+          "temporarily forced to be synchronous instead. "                  \
+          "The default means after an amount of pages proportional to the " \
+          "max capacity is enqueued, we resort to synchronous unmapping.")  \
+                                                                            \
   product(uint, ZStatisticsInterval, 10, DIAGNOSTIC,                        \
           "Time between statistics print outs (in seconds)")                \
           range(1, (uint)-1)                                                \

src/hotspot/share/gc/z/zUnmapper.cpp

@@ -23,6 +23,7 @@
 
 #include "precompiled.hpp"
 #include "gc/shared/gc_globals.hpp"
+#include "gc/shared/gcLogPrecious.hpp"
 #include "gc/z/zList.inline.hpp"
 #include "gc/z/zLock.inline.hpp"
 #include "gc/z/zPage.inline.hpp"
@@ -35,6 +36,8 @@ ZUnmapper::ZUnmapper(ZPageAllocator* page_allocator)
   : _page_allocator(page_allocator),
     _lock(),
     _queue(),
+    _enqueued_bytes(0),
+    _warned_sync_unmapping(false),
     _stop(false) {
   set_name("ZUnmapper");
   create_and_start();
@@ -50,13 +53,45 @@ ZPage* ZUnmapper::dequeue() {
 
     ZPage* const page = _queue.remove_first();
     if (page != nullptr) {
+      _enqueued_bytes -= page->size();
       return page;
     }
 
     _lock.wait();
   }
 }
 
+bool ZUnmapper::try_enqueue(ZPage* page) {
+  // Enqueue for asynchronous unmap and destroy
+  ZLocker<ZConditionLock> locker(&_lock);
+  if (is_saturated()) {
+    // The unmapper thread is lagging behind and is unable to unmap memory fast enough
+    if (!_warned_sync_unmapping) {
+      _warned_sync_unmapping = true;
+      log_warning_p(gc)("WARNING: Encountered synchronous unmapping because asynchronous unmapping could not keep up");
+    }
+    log_debug(gc, unmap)("Synchronous unmapping " SIZE_FORMAT "M page", page->size() / M);
+    return false;
+  }
+
+  log_trace(gc, unmap)("Asynchronous unmapping " SIZE_FORMAT "M page (" SIZE_FORMAT "M / " SIZE_FORMAT "M enqueued)",
+                       page->size() / M, _enqueued_bytes / M, queue_capacity() / M);
+
+  _queue.insert_last(page);
+  _enqueued_bytes += page->size();
+  _lock.notify_all();
+
+  return true;
+}
+
+size_t ZUnmapper::queue_capacity() const {
+  return align_up<size_t>(_page_allocator->max_capacity() * ZAsyncUnmappingLimit / 100.0, ZGranuleSize);
+}
+
+bool ZUnmapper::is_saturated() const {
+  return _enqueued_bytes >= queue_capacity();
+}
+
 void ZUnmapper::do_unmap_and_destroy_page(ZPage* page) const {
   EventZUnmap event;
   const size_t unmapped = page->size();
@@ -70,10 +105,10 @@ void ZUnmapper::do_unmap_and_destroy_page(ZPage* page) const {
 }
 
 void ZUnmapper::unmap_and_destroy_page(ZPage* page) {
-  // Enqueue for asynchronous unmap and destroy
-  ZLocker<ZConditionLock> locker(&_lock);
-  _queue.insert_last(page);
-  _lock.notify_all();
+  if (!try_enqueue(page)) {
+    // Synchronously unmap and destroy
+    do_unmap_and_destroy_page(page);
+  }
 }
 
 void ZUnmapper::run_thread() {

src/hotspot/share/gc/z/zUnmapper.hpp

@@ -36,9 +36,14 @@ class ZUnmapper : public ZThread {
   ZPageAllocator* const _page_allocator;
   ZConditionLock        _lock;
   ZList<ZPage>          _queue;
+  size_t                _enqueued_bytes;
+  bool                  _warned_sync_unmapping;
   bool                  _stop;
 
   ZPage* dequeue();
+  bool try_enqueue(ZPage* page);
+  size_t queue_capacity() const;
+  bool is_saturated() const;
   void do_unmap_and_destroy_page(ZPage* page) const;
 
 protected:

src/hotspot/share/logging/logTag.hpp

@@ -197,6 +197,7 @@ class outputStream;
   LOG_TAG(tlab) \
   LOG_TAG(tracking) \
   LOG_TAG(unload) /* Trace unloading of classes */ \
+  LOG_TAG(unmap) \
   LOG_TAG(unshareable) \
   NOT_PRODUCT(LOG_TAG(upcall)) \
   LOG_TAG(update) \