[scudo] Add primary option to enable/disable cache blocks. (#129794)

When configured this way, no primary blocks will be cached except the
batch class. Nothing else changes, no change in the page releasing
strategy.

Author: cferris1000

compiler-rt/lib/scudo/standalone/CMakeLists.txt

@@ -74,7 +74,6 @@ set(SCUDO_HEADERS
   internal_defs.h
   linux.h
   list.h
-  local_cache.h
   memtag.h
   mem_map.h
   mem_map_base.h
@@ -90,6 +89,7 @@ set(SCUDO_HEADERS
   report.h
   report_linux.h
   secondary.h
+  size_class_allocator.h
   size_class_map.h
   stack_depot.h
   stats.h

compiler-rt/lib/scudo/standalone/allocator_config.def

@@ -81,6 +81,10 @@ PRIMARY_REQUIRED(const s32, MaxReleaseToOsIntervalMs)
 
 // PRIMARY_OPTIONAL(TYPE, NAME, DEFAULT)
 //
+
+// Enables/disables primary block caching. Batch class still caches.
+PRIMARY_OPTIONAL(const bool, EnableBlockCache, true)
+
 // The scale of a compact pointer. E.g., Ptr = Base + (CompactPtr << Scale).
 PRIMARY_OPTIONAL(const uptr, CompactPtrScale, SCUDO_MIN_ALIGNMENT_LOG)
 

compiler-rt/lib/scudo/standalone/combined.h

@@ -15,14 +15,14 @@
 #include "common.h"
 #include "flags.h"
 #include "flags_parser.h"
-#include "local_cache.h"
 #include "mem_map.h"
 #include "memtag.h"
 #include "mutex.h"
 #include "options.h"
 #include "quarantine.h"
 #include "report.h"
 #include "secondary.h"
+#include "size_class_allocator.h"
 #include "stack_depot.h"
 #include "string_utils.h"
 #include "tsd.h"
@@ -54,7 +54,7 @@ class Allocator {
       typename AllocatorConfig::template PrimaryT<PrimaryConfig<Config>>;
   using SecondaryT =
       typename AllocatorConfig::template SecondaryT<SecondaryConfig<Config>>;
-  using CacheT = typename PrimaryT::CacheT;
+  using SizeClassAllocatorT = typename PrimaryT::SizeClassAllocatorT;
   typedef Allocator<Config, PostInitCallback> ThisT;
   typedef typename AllocatorConfig::template TSDRegistryT<ThisT> TSDRegistryT;
 
@@ -63,8 +63,9 @@ class Allocator {
   }
 
   struct QuarantineCallback {
-    explicit QuarantineCallback(ThisT &Instance, CacheT &LocalCache)
-        : Allocator(Instance), Cache(LocalCache) {}
+    explicit QuarantineCallback(ThisT &Instance,
+                                SizeClassAllocatorT &SizeClassAllocator)
+        : Allocator(Instance), SizeClassAllocator(SizeClassAllocator) {}
 
     // Chunk recycling function, returns a quarantined chunk to the backend,
     // first making sure it hasn't been tampered with.
@@ -80,7 +81,7 @@ class Allocator {
       if (allocatorSupportsMemoryTagging<AllocatorConfig>())
         Ptr = untagPointer(Ptr);
       void *BlockBegin = Allocator::getBlockBegin(Ptr, &Header);
-      Cache.deallocate(Header.ClassId, BlockBegin);
+      SizeClassAllocator.deallocate(Header.ClassId, BlockBegin);
     }
 
     // We take a shortcut when allocating a quarantine batch by working with the
@@ -89,7 +90,7 @@ class Allocator {
     void *allocate(UNUSED uptr Size) {
       const uptr QuarantineClassId = SizeClassMap::getClassIdBySize(
           sizeof(QuarantineBatch) + Chunk::getHeaderSize());
-      void *Ptr = Cache.allocate(QuarantineClassId);
+      void *Ptr = SizeClassAllocator.allocate(QuarantineClassId);
       // Quarantine batch allocation failure is fatal.
       if (UNLIKELY(!Ptr))
         reportOutOfMemory(SizeClassMap::getSizeByClassId(QuarantineClassId));
@@ -126,14 +127,15 @@ class Allocator {
 
       Header.State = Chunk::State::Available;
       Chunk::storeHeader(Allocator.Cookie, Ptr, &Header);
-      Cache.deallocate(QuarantineClassId,
-                       reinterpret_cast<void *>(reinterpret_cast<uptr>(Ptr) -
-                                                Chunk::getHeaderSize()));
+      SizeClassAllocator.deallocate(
+          QuarantineClassId,
+          reinterpret_cast<void *>(reinterpret_cast<uptr>(Ptr) -
+                                   Chunk::getHeaderSize()));
     }
 
   private:
     ThisT &Allocator;
-    CacheT &Cache;
+    SizeClassAllocatorT &SizeClassAllocator;
   };
 
   typedef GlobalQuarantine<QuarantineCallback, void> QuarantineT;
@@ -263,7 +265,9 @@ class Allocator {
   QuarantineT *getQuarantine() { return &Quarantine; }
 
   // The Cache must be provided zero-initialized.
-  void initCache(CacheT *Cache) { Cache->init(&Stats, &Primary); }
+  void initAllocator(SizeClassAllocatorT *SizeClassAllocator) {
+    SizeClassAllocator->init(&Stats, &Primary);
+  }
 
   // Release the resources used by a TSD, which involves:
   // - draining the local quarantine cache to the global quarantine;
@@ -273,15 +277,16 @@ class Allocator {
   void commitBack(TSD<ThisT> *TSD) {
     TSD->assertLocked(/*BypassCheck=*/true);
     Quarantine.drain(&TSD->getQuarantineCache(),
-                     QuarantineCallback(*this, TSD->getCache()));
-    TSD->getCache().destroy(&Stats);
+                     QuarantineCallback(*this, TSD->getSizeClassAllocator()));
+    TSD->getSizeClassAllocator().destroy(&Stats);
   }
 
   void drainCache(TSD<ThisT> *TSD) {
     TSD->assertLocked(/*BypassCheck=*/true);
-    Quarantine.drainAndRecycle(&TSD->getQuarantineCache(),
-                               QuarantineCallback(*this, TSD->getCache()));
-    TSD->getCache().drain();
+    Quarantine.drainAndRecycle(
+        &TSD->getQuarantineCache(),
+        QuarantineCallback(*this, TSD->getSizeClassAllocator()));
+    TSD->getSizeClassAllocator().drain();
   }
   void drainCaches() { TSDRegistry.drainCaches(this); }
 
@@ -390,13 +395,13 @@ class Allocator {
       ClassId = SizeClassMap::getClassIdBySize(NeededSize);
       DCHECK_NE(ClassId, 0U);
       typename TSDRegistryT::ScopedTSD TSD(TSDRegistry);
-      Block = TSD->getCache().allocate(ClassId);
+      Block = TSD->getSizeClassAllocator().allocate(ClassId);
       // If the allocation failed, retry in each successively larger class until
       // it fits. If it fails to fit in the largest class, fallback to the
       // Secondary.
       if (UNLIKELY(!Block)) {
         while (ClassId < SizeClassMap::LargestClassId && !Block)
-          Block = TSD->getCache().allocate(++ClassId);
+          Block = TSD->getSizeClassAllocator().allocate(++ClassId);
         if (!Block)
           ClassId = 0;
       }
@@ -1280,7 +1285,8 @@ class Allocator {
         bool CacheDrained;
         {
           typename TSDRegistryT::ScopedTSD TSD(TSDRegistry);
-          CacheDrained = TSD->getCache().deallocate(ClassId, BlockBegin);
+          CacheDrained =
+              TSD->getSizeClassAllocator().deallocate(ClassId, BlockBegin);
         }
         // When we have drained some blocks back to the Primary from TSD, that
         // implies that we may have the chance to release some pages as well.
@@ -1296,7 +1302,8 @@ class Allocator {
         retagBlock(Options, TaggedPtr, Ptr, Header, Size, false);
       typename TSDRegistryT::ScopedTSD TSD(TSDRegistry);
       Quarantine.put(&TSD->getQuarantineCache(),
-                     QuarantineCallback(*this, TSD->getCache()), Ptr, Size);
+                     QuarantineCallback(*this, TSD->getSizeClassAllocator()),
+                     Ptr, Size);
     }
   }
 

compiler-rt/lib/scudo/standalone/primary32.h

@@ -13,10 +13,10 @@
 #include "bytemap.h"
 #include "common.h"
 #include "list.h"
-#include "local_cache.h"
 #include "options.h"
 #include "release.h"
 #include "report.h"
+#include "size_class_allocator.h"
 #include "stats.h"
 #include "string_utils.h"
 #include "thread_annotations.h"
@@ -52,7 +52,10 @@ template <typename Config> class SizeClassAllocator32 {
   static_assert((1UL << Config::getRegionSizeLog()) >= SizeClassMap::MaxSize,
                 "");
   typedef SizeClassAllocator32<Config> ThisT;
-  typedef SizeClassAllocatorLocalCache<ThisT> CacheT;
+  using SizeClassAllocatorT =
+      typename Conditional<Config::getEnableBlockCache(),
+                           SizeClassAllocatorLocalCache<ThisT>,
+                           SizeClassAllocatorNoCache<ThisT>>::type;
   typedef TransferBatch<ThisT> TransferBatchT;
   typedef BatchGroup<ThisT> BatchGroupT;
 
@@ -191,25 +194,28 @@ template <typename Config> class SizeClassAllocator32 {
     return BlockSize > PageSize;
   }
 
-  u16 popBlocks(CacheT *C, uptr ClassId, CompactPtrT *ToArray,
-                const u16 MaxBlockCount) {
+  u16 popBlocks(SizeClassAllocatorT *SizeClassAllocator, uptr ClassId,
+                CompactPtrT *ToArray, const u16 MaxBlockCount) {
     DCHECK_LT(ClassId, NumClasses);
     SizeClassInfo *Sci = getSizeClassInfo(ClassId);
     ScopedLock L(Sci->Mutex);
 
-    u16 PopCount = popBlocksImpl(C, ClassId, Sci, ToArray, MaxBlockCount);
+    u16 PopCount =
+        popBlocksImpl(SizeClassAllocator, ClassId, Sci, ToArray, MaxBlockCount);
     if (UNLIKELY(PopCount == 0)) {
-      if (UNLIKELY(!populateFreeList(C, ClassId, Sci)))
+      if (UNLIKELY(!populateFreeList(SizeClassAllocator, ClassId, Sci)))
         return 0U;
-      PopCount = popBlocksImpl(C, ClassId, Sci, ToArray, MaxBlockCount);
+      PopCount = popBlocksImpl(SizeClassAllocator, ClassId, Sci, ToArray,
+                               MaxBlockCount);
       DCHECK_NE(PopCount, 0U);
     }
 
     return PopCount;
   }
 
   // Push the array of free blocks to the designated batch group.
-  void pushBlocks(CacheT *C, uptr ClassId, CompactPtrT *Array, u32 Size) {
+  void pushBlocks(SizeClassAllocatorT *SizeClassAllocator, uptr ClassId,
+                  CompactPtrT *Array, u32 Size) {
     DCHECK_LT(ClassId, NumClasses);
     DCHECK_GT(Size, 0);
 
@@ -240,7 +246,7 @@ template <typename Config> class SizeClassAllocator32 {
     }
 
     ScopedLock L(Sci->Mutex);
-    pushBlocksImpl(C, ClassId, Sci, Array, Size, SameGroup);
+    pushBlocksImpl(SizeClassAllocator, ClassId, Sci, Array, Size, SameGroup);
   }
 
   void disable() NO_THREAD_SAFETY_ANALYSIS {
@@ -529,8 +535,8 @@ template <typename Config> class SizeClassAllocator32 {
       // memory group here.
       BG->CompactPtrGroupBase = 0;
       BG->BytesInBGAtLastCheckpoint = 0;
-      BG->MaxCachedPerBatch =
-          CacheT::getMaxCached(getSizeByClassId(SizeClassMap::BatchClassId));
+      BG->MaxCachedPerBatch = SizeClassAllocatorT::getMaxCached(
+          getSizeByClassId(SizeClassMap::BatchClassId));
 
       Sci->FreeListInfo.BlockList.push_front(BG);
     }
@@ -597,18 +603,18 @@ template <typename Config> class SizeClassAllocator32 {
   // same group then we will skip checking the group id of each block.
   //
   // The region mutex needs to be held while calling this method.
-  void pushBlocksImpl(CacheT *C, uptr ClassId, SizeClassInfo *Sci,
-                      CompactPtrT *Array, u32 Size, bool SameGroup = false)
-      REQUIRES(Sci->Mutex) {
+  void pushBlocksImpl(SizeClassAllocatorT *SizeClassAllocator, uptr ClassId,
+                      SizeClassInfo *Sci, CompactPtrT *Array, u32 Size,
+                      bool SameGroup = false) REQUIRES(Sci->Mutex) {
     DCHECK_NE(ClassId, SizeClassMap::BatchClassId);
     DCHECK_GT(Size, 0U);
 
     auto CreateGroup = [&](uptr CompactPtrGroupBase) {
-      BatchGroupT *BG =
-          reinterpret_cast<BatchGroupT *>(C->getBatchClassBlock());
+      BatchGroupT *BG = reinterpret_cast<BatchGroupT *>(
+          SizeClassAllocator->getBatchClassBlock());
       BG->Batches.clear();
-      TransferBatchT *TB =
-          reinterpret_cast<TransferBatchT *>(C->getBatchClassBlock());
+      TransferBatchT *TB = reinterpret_cast<TransferBatchT *>(
+          SizeClassAllocator->getBatchClassBlock());
       TB->clear();
 
       BG->CompactPtrGroupBase = CompactPtrGroupBase;
@@ -629,8 +635,8 @@ template <typename Config> class SizeClassAllocator32 {
         u16 UnusedSlots =
             static_cast<u16>(BG->MaxCachedPerBatch - CurBatch->getCount());
         if (UnusedSlots == 0) {
-          CurBatch =
-              reinterpret_cast<TransferBatchT *>(C->getBatchClassBlock());
+          CurBatch = reinterpret_cast<TransferBatchT *>(
+              SizeClassAllocator->getBatchClassBlock());
           CurBatch->clear();
           Batches.push_front(CurBatch);
           UnusedSlots = BG->MaxCachedPerBatch;
@@ -704,9 +710,9 @@ template <typename Config> class SizeClassAllocator32 {
     InsertBlocks(Cur, Array + Size - Count, Count);
   }
 
-  u16 popBlocksImpl(CacheT *C, uptr ClassId, SizeClassInfo *Sci,
-                    CompactPtrT *ToArray, const u16 MaxBlockCount)
-      REQUIRES(Sci->Mutex) {
+  u16 popBlocksImpl(SizeClassAllocatorT *SizeClassAllocator, uptr ClassId,
+                    SizeClassInfo *Sci, CompactPtrT *ToArray,
+                    const u16 MaxBlockCount) REQUIRES(Sci->Mutex) {
     if (Sci->FreeListInfo.BlockList.empty())
       return 0U;
 
@@ -730,11 +736,11 @@ template <typename Config> class SizeClassAllocator32 {
     // So far, instead of always filling the blocks to `MaxBlockCount`, we only
     // examine single `TransferBatch` to minimize the time spent on the primary
     // allocator. Besides, the sizes of `TransferBatch` and
-    // `CacheT::getMaxCached()` may also impact the time spent on accessing the
-    // primary allocator.
+    // `SizeClassAllocatorT::getMaxCached()` may also impact the time spent on
+    // accessing the primary allocator.
     // TODO(chiahungduan): Evaluate if we want to always prepare `MaxBlockCount`
     // blocks and/or adjust the size of `TransferBatch` according to
-    // `CacheT::getMaxCached()`.
+    // `SizeClassAllocatorT::getMaxCached()`.
     TransferBatchT *B = Batches.front();
     DCHECK_NE(B, nullptr);
     DCHECK_GT(B->getCount(), 0U);
@@ -754,7 +760,7 @@ template <typename Config> class SizeClassAllocator32 {
       // deallocate. Read the comment in `pushBatchClassBlocks()` for more
       // details.
       if (ClassId != SizeClassMap::BatchClassId)
-        C->deallocate(SizeClassMap::BatchClassId, B);
+        SizeClassAllocator->deallocate(SizeClassMap::BatchClassId, B);
 
       if (Batches.empty()) {
         BatchGroupT *BG = Sci->FreeListInfo.BlockList.front();
@@ -766,15 +772,16 @@ template <typename Config> class SizeClassAllocator32 {
         // Which means, once we pop the last TransferBatch, the block is
         // implicitly deallocated.
         if (ClassId != SizeClassMap::BatchClassId)
-          C->deallocate(SizeClassMap::BatchClassId, BG);
+          SizeClassAllocator->deallocate(SizeClassMap::BatchClassId, BG);
       }
     }
 
     Sci->FreeListInfo.PoppedBlocks += PopCount;
     return PopCount;
   }
 
-  NOINLINE bool populateFreeList(CacheT *C, uptr ClassId, SizeClassInfo *Sci)
+  NOINLINE bool populateFreeList(SizeClassAllocatorT *SizeClassAllocator,
+                                 uptr ClassId, SizeClassInfo *Sci)
       REQUIRES(Sci->Mutex) {
     uptr Region;
     uptr Offset;
@@ -791,13 +798,13 @@ template <typename Config> class SizeClassAllocator32 {
       Region = allocateRegion(Sci, ClassId);
       if (UNLIKELY(!Region))
         return false;
-      C->getStats().add(StatMapped, RegionSize);
+      SizeClassAllocator->getStats().add(StatMapped, RegionSize);
       Sci->CurrentRegion = Region;
       Offset = 0;
     }
 
     const uptr Size = getSizeByClassId(ClassId);
-    const u16 MaxCount = CacheT::getMaxCached(Size);
+    const u16 MaxCount = SizeClassAllocatorT::getMaxCached(Size);
     DCHECK_GT(MaxCount, 0U);
     // The maximum number of blocks we should carve in the region is dictated
     // by the maximum number of batches we want to fill, and the amount of
@@ -827,7 +834,8 @@ template <typename Config> class SizeClassAllocator32 {
       for (u32 I = 1; I < NumberOfBlocks; I++) {
         if (UNLIKELY(compactPtrGroupBase(ShuffleArray[I]) != CurGroup)) {
           shuffle(ShuffleArray + I - N, N, &Sci->RandState);
-          pushBlocksImpl(C, ClassId, Sci, ShuffleArray + I - N, N,
+          pushBlocksImpl(SizeClassAllocator, ClassId, Sci, ShuffleArray + I - N,
+                         N,
                          /*SameGroup=*/true);
           N = 1;
           CurGroup = compactPtrGroupBase(ShuffleArray[I]);
@@ -837,7 +845,8 @@ template <typename Config> class SizeClassAllocator32 {
       }
 
       shuffle(ShuffleArray + NumberOfBlocks - N, N, &Sci->RandState);
-      pushBlocksImpl(C, ClassId, Sci, &ShuffleArray[NumberOfBlocks - N], N,
+      pushBlocksImpl(SizeClassAllocator, ClassId, Sci,
+                     &ShuffleArray[NumberOfBlocks - N], N,
                      /*SameGroup=*/true);
     } else {
       pushBatchClassBlocks(Sci, ShuffleArray, NumberOfBlocks);
@@ -850,7 +859,7 @@ template <typename Config> class SizeClassAllocator32 {
     Sci->FreeListInfo.PushedBlocks -= NumberOfBlocks;
 
     const uptr AllocatedUser = Size * NumberOfBlocks;
-    C->getStats().add(StatFree, AllocatedUser);
+    SizeClassAllocator->getStats().add(StatFree, AllocatedUser);
     DCHECK_LE(Sci->CurrentRegionAllocated + AllocatedUser, RegionSize);
     // If there is not enough room in the region currently associated to fit
     // more blocks, we deassociate the region by resetting CurrentRegion and