Report DOM partitionAlloc size to V8 GC.

Source/wtf/PartitionAlloc.cpp

Index: Source/wtf/PartitionAlloc.cpp
diff --git a/Source/wtf/PartitionAlloc.cpp b/Source/wtf/PartitionAlloc.cpp
index a63a684c51980189aed9ab281b326c9f54d06ac4..f4fbdea72ba879b56c125489fad8f68134534a00 100644
--- a/Source/wtf/PartitionAlloc.cpp
+++ b/Source/wtf/PartitionAlloc.cpp
@@ -112,6 +112,7 @@ static void parititonAllocBaseInit(PartitionRootBase* root)
     spinLockUnlock(&PartitionRootBase::gInitializedLock);
 
     root->initialized = true;
+    root->totalSizeOfNonFreePages = 0;

[Chris Evans, 2014/06/06 18:30:50]

I think we should name this totalSizeOfCommittedPages, it's clearer what it is.

[kouhei (in TOK), 2014/06/09 05:48:22]

Done.

     root->totalSizeOfSuperPages = 0;
     root->nextSuperPage = 0;
     root->nextPartitionPage = 0;
@@ -314,6 +315,7 @@ static ALWAYS_INLINE void* partitionAllocPartitionPages(PartitionRootBase* root,
     ASSERT(!(reinterpret_cast<uintptr_t>(root->nextPartitionPageEnd) % kPartitionPageSize));
     RELEASE_ASSERT(numPartitionPages <= kNumPartitionPagesPerSuperPage);
     size_t totalSize = kPartitionPageSize * numPartitionPages;
+    root->totalSizeOfNonFreePages += totalSize;
     size_t numPartitionPagesLeft = (root->nextPartitionPageEnd - root->nextPartitionPage) >> kPartitionPageShift;
     if (LIKELY(numPartitionPagesLeft >= numPartitionPages)) {
         // In this case, we can still hand out pages from the current super page
@@ -675,7 +677,9 @@ void* partitionAllocSlowPath(PartitionRootBase* root, int flags, size_t size, Pa
         ASSERT(newPage->freeCacheIndex == -1);
         bucket->freePagesHead = newPage->nextPage;
         void* addr = partitionPageToPointer(newPage);
-        recommitSystemPages(addr, newPage->bucket->numSystemPagesPerSlotSpan * kSystemPageSize);
+        size_t commitSize = newPage->bucket->numSystemPagesPerSlotSpan * kSystemPageSize;
+        recommitSystemPages(addr, commitSize);
+        root->totalSizeOfNonFreePages += commitSize;

[Chris Evans, 2014/06/06 18:30:50]

I think this is the right model: adjusting the 

However, we've missed the adjustment in the other call to recommitSystemPages in
this file. I'd be ok with doing this more safely: do that actual size adjustment
in a helper function that does the adjustment and then calls
recommitSystemPages. We could then re-use that code. If it means the need to
pass the root around more, I'm fine with that.

Also, we've missed an adjustment around another decommitSystemPages call. We
need to get them all :-) Same solution applies.

[kouhei (in TOK), 2014/06/09 05:48:22]

Done. I added partition{Re,De}commitPages wrapper func.

     } else {
         // Third. If we get here, we need a brand new page.
         size_t numPartitionPages = partitionBucketPartitionPages(bucket);
@@ -710,6 +714,10 @@ static ALWAYS_INLINE void partitionFreePage(PartitionPage* page)
 static ALWAYS_INLINE void partitionRegisterEmptyPage(PartitionPage* page)
 {
     PartitionRootBase* root = partitionPageToRoot(page);
+    size_t pageSize = page->bucket->numSystemPagesPerSlotSpan * kSystemPageSize;
+    ASSERT(root->totalSizeOfNonFreePages > pageSize);
+    root->totalSizeOfNonFreePages -= pageSize;

[Chris Evans, 2014/06/06 18:30:50]

I do not think this is the correct place to do this. The page is empty but is
still committed. I think we might do it in a wrapper around the call to
decommitSystemPages(), same idea as above. Again, this would capture both places
we need to do it.

[kouhei (in TOK), 2014/06/09 05:48:22]

Removed this, and moved counter decrement into partitionDecommitSystemPages().

+
     // If the page is already registered as empty, give it another life.
     if (page->freeCacheIndex != -1) {
         ASSERT(page->freeCacheIndex >= 0);
@@ -757,7 +765,7 @@ void partitionFreeSlowPath(PartitionPage* page)
             partitionDirectUnmap(page);
             return;
         }
-        // If it's the current page, attempt to change it. We'd prefer to leave
+        // If it's the current active page, attempt to change it. We'd prefer to leave
         // the page empty as a gentle force towards defragmentation.
         if (LIKELY(page == bucket->activePagesHead) && page->nextPage) {
             if (partitionSetNewActivePage(page->nextPage)) {