fix race in moveRegularItemWith sync where insertOrReplace can cause move to fail

Author: byrnedj

cachelib/allocator/CacheAllocator-inl.h

@@ -1294,8 +1294,21 @@ size_t CacheAllocator<CacheTrait>::wakeUpWaitersLocked(folly::StringPiece key,
 }
 
 template <typename CacheTrait>
-void CacheAllocator<CacheTrait>::moveRegularItemWithSync(
+bool CacheAllocator<CacheTrait>::moveRegularItemWithSync(
     Item& oldItem, WriteHandle& newItemHdl) {
+  //on function exit - the new item handle is no longer moving
+  //and other threads may access it - but in case where
+  //we failed to replace in access container we can give the
+  //new item back to the allocator
+  auto guard = folly::makeGuard([&]() {
+    auto ref = newItemHdl->unmarkMoving();
+    if (UNLIKELY(ref == 0)) {
+      const auto res =
+          releaseBackToAllocator(*newItemHdl, RemoveContext::kNormal, false);
+      XDCHECK(res == ReleaseRes::kReleased);
+    }
+  });
+
   XDCHECK(oldItem.isMoving());
   XDCHECK(!oldItem.isExpired());
   // TODO: should we introduce new latency tracker. E.g. evictRegularLatency_
@@ -1326,6 +1339,22 @@ void CacheAllocator<CacheTrait>::moveRegularItemWithSync(
 
   auto replaced = accessContainer_->replaceIf(oldItem, *newItemHdl,
                                    predicate);
+  // another thread may have called insertOrReplace which could have
+  // marked this item as unaccessible causing the replaceIf
+  // in the access container to fail - in this case we want
+  // to abort the move since the item is no longer valid
+  if (!replaced) {
+      return false;
+  }
+  // what if another thread calls insertOrReplace now when
+  // the item is moving and already replaced in the hash table?
+  // 1. it succeeds in updating the hash table - so there is
+  //    no guarentee that isAccessible() is true
+  // 2. it will then try to remove from MM container
+  //     - this operation will wait for newItemHdl to
+  //       be unmarkedMoving via the waitContext
+  // 3. replaced handle is returned and eventually drops
+  //    ref to 0 and the item is recycled back to allocator.
 
   if (config_.moveCb) {
     // Execute the move callback. We cannot make any guarantees about the
@@ -1367,14 +1396,7 @@ void CacheAllocator<CacheTrait>::moveRegularItemWithSync(
     XDCHECK(newItemHdl->hasChainedItem());
   }
   newItemHdl.unmarkNascent();
-  auto ref = newItemHdl->unmarkMoving();
-  //remove because there is a chance the new item was not
-  //added to the access container
-  if (UNLIKELY(ref == 0)) {
-    const auto res =
-        releaseBackToAllocator(*newItemHdl, RemoveContext::kNormal, false);
-    XDCHECK(res == ReleaseRes::kReleased);
-  }
+  return true;
 }
 
 template <typename CacheTrait>
@@ -1756,7 +1778,10 @@ CacheAllocator<CacheTrait>::tryEvictToNextMemoryTier(
 
     if (newItemHdl) {
       XDCHECK_EQ(newItemHdl->getSize(), item.getSize());
-      moveRegularItemWithSync(item, newItemHdl);
+      if (!moveRegularItemWithSync(item, newItemHdl)) {
+          return WriteHandle{};
+      }
+      XDCHECK_EQ(newItemHdl->getKey(),item.getKey());
       item.unmarkMoving();
       return newItemHdl;
     } else {
@@ -1795,7 +1820,9 @@ CacheAllocator<CacheTrait>::tryPromoteToNextMemoryTier(
 
     if (newItemHdl) {
       XDCHECK_EQ(newItemHdl->getSize(), item.getSize());
-      moveRegularItemWithSync(item, newItemHdl);
+      if (!moveRegularItemWithSync(item, newItemHdl)) {
+          return WriteHandle{};
+      }
       item.unmarkMoving();
       return newItemHdl;
     } else {
@@ -3148,9 +3175,13 @@ bool CacheAllocator<CacheTrait>::tryMovingForSlabRelease(
       // TODO: add support for chained items
       return false;
     } else {
-      moveRegularItemWithSync(oldItem, newItemHdl);
+      //move can fail if another thread calls insertOrReplace
+      //in this case oldItem is no longer valid (not accessible, 
+      //it gets removed from MMContainer and evictForSlabRelease
+      //will send it back to the allocator
+      bool ret = moveRegularItemWithSync(oldItem, newItemHdl);
       removeFromMMContainer(oldItem);
-      return true;
+      return ret;
     }
   }
 }

cachelib/allocator/CacheAllocator.h

@@ -1617,7 +1617,7 @@ class CacheAllocator : public CacheBase {
   //
   // @return true  If the move was completed, and the containers were updated
   //               successfully.
-  void moveRegularItemWithSync(Item& oldItem, WriteHandle& newItemHdl);
+  bool moveRegularItemWithSync(Item& oldItem, WriteHandle& newItemHdl);
 
   // Moves a regular item to a different slab. This should only be used during
   // slab release after the item's exclusive bit has been set. The user supplied