Add Decay Range.

Author: mjp41

src/backend/backend.h

@@ -3,6 +3,7 @@
 #include "../pal/pal.h"
 #include "commitrange.h"
 #include "commonconfig.h"
+#include "decayrange.h"
 #include "empty_range.h"
 #include "globalrange.h"
 #include "largebuddyrange.h"
@@ -148,9 +149,10 @@ namespace snmalloc
     using GlobalR = GlobalRange<StatsR>;
 
 #  ifdef SNMALLOC_META_PROTECTED
+    using CommittedRange =
+      DecayRange<CommitRange<GlobalR, DefaultPal>, DefaultPal, Pagemap>;
     // Source for object allocations
-    using ObjectRange =
-      LargeBuddyRange<CommitRange<GlobalR, DefaultPal>, 21, 21, Pagemap>;
+    using ObjectRange = LargeBuddyRange<CommittedRange, 21, 21, Pagemap>;
     // Set up protected range for metadata
     using SubR = CommitRange<SubRange<GlobalR, DefaultPal, 6>, DefaultPal>;
     using MetaRange =
@@ -159,8 +161,10 @@ namespace snmalloc
 #  else
     // Source for object allocations and metadata
     // No separation between the two
-    using ObjectRange = SmallBuddyRange<
-      LargeBuddyRange<CommitRange<GlobalR, DefaultPal>, 21, 21, Pagemap>>;
+    using CommittedRange =
+      DecayRange<CommitRange<GlobalR, DefaultPal>, DefaultPal, Pagemap>;
+    using ObjectRange =
+      SmallBuddyRange<LargeBuddyRange<CommittedRange, 21, 21, Pagemap>>;
     using GlobalMetaRange = GlobalRange<ObjectRange>;
 #  endif
 #endif

src/backend/decayrange.h

@@ -0,0 +1,351 @@
+#pragma once
+
+#include "../ds/ptrwrap.h"
+#include "../pal/pal_ds.h"
+#include "largebuddyrange.h"
+
+namespace snmalloc
+{
+  template<SNMALLOC_CONCEPT(RBRep) Rep>
+  class RepList
+  {
+    uintptr_t head = 0;
+
+    RepList(uintptr_t head) : head(head) {}
+
+  public:
+    constexpr RepList() = default;
+
+    [[nodiscard]] bool is_empty() const
+    {
+      return head == 0;
+    }
+
+    RepList get_next()
+    {
+      SNMALLOC_ASSERT(!is_empty());
+      auto next_field = &(Rep::ref(false, head));
+      auto next = Rep::get(next_field);
+      return {next};
+    }
+
+    capptr::Chunk<void> get_capability()
+    {
+      return capptr::Chunk<void>(reinterpret_cast<void*>(head));
+    }
+
+    RepList cons(capptr::Chunk<void> new_head_cap)
+    {
+      auto new_head = new_head_cap.unsafe_uintptr();
+      auto field = &(Rep::ref(false, new_head));
+      Rep::set(field, head);
+      return {new_head};
+    }
+
+    template<typename F>
+    void forall(F f)
+    {
+      auto curr = *this;
+      while (!curr.is_empty())
+      {
+        auto next = curr.get_next();
+
+        f(curr.get_capability());
+
+        curr = next;
+      }
+    }
+  };
+
+  /**
+   * Concurrent Stack
+   *
+   * This stack supports the following clients
+   * (push|pop)* || pop_all* || ... || pop_all*
+   *
+   * That is a single thread that can do push and pop, and other threads
+   * that do pop_all.  pop_all if it returns a value, returns all of the
+   * stack, however, it may return nullptr if it races with either a push
+   * or a pop.
+   *
+   * The primary use case is single-threaded access, where other threads
+   * can attempt to steal all the values.
+   */
+  template<SNMALLOC_CONCEPT(RBRep) Rep>
+  class RepStack
+  {
+    static constexpr auto empty = RepList<Rep>{};
+
+  private:
+    alignas(CACHELINE_SIZE) std::atomic<RepList<Rep>> stack{};
+
+    RepList<Rep> take()
+    {
+      if (stack.load(std::memory_order_relaxed).is_empty())
+        return empty;
+      return stack.exchange(empty, std::memory_order_acquire);
+    }
+
+    void replace(RepList<Rep> new_head)
+    {
+      SNMALLOC_ASSERT(stack.load().is_empty());
+      stack.store(new_head, std::memory_order_release);
+    }
+
+  public:
+    constexpr RepStack() = default;
+
+    void push(capptr::Chunk<void> new_head_cap)
+    {
+      auto old_head = take();
+      auto new_head = old_head.cons(new_head_cap);
+      replace(new_head);
+    }
+
+    capptr::Chunk<void> pop()
+    {
+      auto old_head = take();
+      if (old_head.is_empty())
+        return nullptr;
+
+      auto next = old_head.get_next();
+      replace(next);
+
+      return old_head.get_capability();
+    }
+
+    RepList<Rep> pop_all()
+    {
+      return take();
+    }
+  };
+
+  /**
+   * This range slowly filters back memory to the parent range.
+   * It locally caches memory and after it hasn't been used for some time
+   * it goes back to its parent range.
+   */
+
+  template<typename ParentRange, typename PAL, typename Pagemap>
+  class DecayRange
+  {
+    /**
+     * How many slab sizes that can be provided.
+     */
+    static constexpr size_t NUM_SLAB_SIZES = Pal::address_bits - MIN_CHUNK_BITS;
+
+    /**
+     * Number of free stacks per chunk size that each allocator will use.
+     * For performance ideally a power of 2.  We will return to the central
+     * pool anything that has not be used in the last NUM_EPOCHS - 1, where
+     * each epoch is separated by DecayMemoryTimerObject::PERIOD.
+     * I.e. if period is 500ms and num of epochs is 4, then we will return to
+     * the central pool anything not used for the last 1500-2000ms.
+     */
+    static constexpr size_t NUM_EPOCHS = 4;
+    static_assert(bits::is_pow2(NUM_EPOCHS), "Code assumes power of two.");
+
+    /**
+     * Stack of ranges that have been returned for reuse.
+     */
+    ModArray<
+      NUM_SLAB_SIZES,
+      ModArray<NUM_EPOCHS, RepStack<BuddyChunkRep<Pagemap>>>>
+      chunk_stack;
+
+    typename ParentRange::State parent{};
+
+    /**
+     * Which is the current epoch to place dealloced chunks, and the
+     * first place we look for allocating chunks.
+     */
+    static inline // alignas(CACHELINE_SIZE)
+      std::atomic<size_t>
+        epoch{0};
+
+    /**
+     * Flag to ensure one-shot registration with the PAL.
+     */
+    static inline std::atomic_bool registered_timer{false};
+
+    std::atomic_bool registered_local{false};
+
+    /**
+     * All activated DecayRanges.
+     */
+    static inline std::atomic<DecayRange*> all_local{nullptr};
+
+    DecayRange* all_local_next{nullptr};
+
+    static void handle_decay_tick()
+    {
+      static_assert(
+        ParentRange::ConcurrencySafe,
+        "Parent must be concurrency safe, as dealloc_range is called here on "
+        "potentially another thread's state.");
+      auto new_epoch = (epoch + 1) % NUM_EPOCHS;
+      // Flush old index for all threads.
+      auto curr = all_local.load(std::memory_order_acquire);
+      while (curr != nullptr)
+      {
+        for (size_t sc = 0; sc < NUM_SLAB_SIZES; sc++)
+        {
+          auto old_stack = curr->chunk_stack[sc][new_epoch].pop_all();
+
+          old_stack.forall([curr, sc](auto cap) {
+            curr->parent->dealloc_range(cap, MIN_CHUNK_SIZE << sc);
+          });
+        }
+        curr = curr->all_local_next;
+      }
+
+      // Advance current index
+      epoch = new_epoch;
+    }
+
+    class DecayMemoryTimerObject : public PalTimerObject
+    {
+      /***
+       * Method for callback object to perform lazy decommit.
+       */
+      static void process(PalTimerObject*)
+      {
+#ifdef SNMALLOC_TRACING
+        message<1024>("DecayRange::handle_decay_tick timer");
+#endif
+        handle_decay_tick();
+      }
+
+      // Specify that we notify the ChunkAllocator every 500ms.
+      static constexpr size_t PERIOD = 500;
+
+    public:
+      constexpr DecayMemoryTimerObject() : PalTimerObject(&process, PERIOD) {}
+    };
+
+    static inline DecayMemoryTimerObject timer_object;
+
+  public:
+    class State
+    {
+      DecayRange commit_range{};
+
+    public:
+      constexpr State() = default;
+
+      DecayRange* operator->()
+      {
+        return &commit_range;
+      }
+    };
+
+    static constexpr bool Aligned = ParentRange::Aligned;
+
+    static constexpr bool ConcurrencySafe = false;
+
+    constexpr DecayRange() = default;
+
+    capptr::Chunk<void> alloc_range(size_t size)
+    {
+      // Check local cache
+
+      if constexpr (pal_supports<Time, PAL>)
+      {
+        auto slab_sizeclass = bits::next_pow2_bits(size) - MIN_CHUNK_BITS;
+        // Try local cache of chunks first
+        for (size_t e = 0; e < NUM_EPOCHS; e++)
+        {
+          auto p = chunk_stack[slab_sizeclass][(epoch - e) % NUM_EPOCHS].pop();
+
+          if (p != nullptr)
+          {
+#ifdef SNMALLOC_TRACING
+            message<1024>(
+              "DecayRange::alloc_range: returning from local cache: {} on {}",
+              address_cast(p),
+              this);
+#endif
+            return p;
+          }
+        }
+      }
+
+      // Loop to possibly flush all the other local threads caches.
+      // Note that flushing passes to the parent range, which may consolidate
+      // blocks and thus be able to service this request.
+      // Alternatively, we could implement stealing, but that wouldn't
+      // be able to consolidate.
+      capptr::Chunk<void> result;
+      for (auto i = NUM_EPOCHS; i > 0; i--)
+      {
+        // Nothing in local cache, so allocate from parent.
+        result = parent->alloc_range(size);
+        if (result != nullptr)
+        {
+#ifdef SNMALLOC_TRACING
+          message<1024>(
+            "DecayRange::alloc_range: returning from parent: {} on {}",
+            address_cast(result),
+            this);
+#endif
+          return result;
+        }
+
+        // We have run out of memory.
+        // Try to free some memory to the parent.
+#ifdef SNMALLOC_TRACING
+        message<1024>("DecayRange::handle_decay_tick OOM");
+#endif
+        handle_decay_tick();
+      }
+
+      // Last try.
+      result = parent->alloc_range(size);
+
+#ifdef SNMALLOC_TRACING
+      message<1024>(
+        "DecayRange::alloc_range: returning from parent last try: {} on {}",
+        address_cast(result),
+        this);
+#endif
+
+      return result;
+    }
+
+    void dealloc_range(capptr::Chunk<void> base, size_t size)
+    {
+      if constexpr (!pal_supports<Time, PAL>)
+      {
+        parent->dealloc_range(base, size);
+        return;
+      }
+
+      if (!registered_timer.exchange(true))
+      {
+        // Register with the PAL.
+        PAL::register_timer(&timer_object);
+      }
+
+      // Check we have registered
+      if (!registered_local.exchange(true))
+      {
+        // Add to the list of local states.
+        auto* head = all_local.load();
+        do
+        {
+          all_local_next = head;
+        } while (!all_local.compare_exchange_strong(head, this));
+      }
+
+      auto slab_sizeclass = bits::next_pow2_bits(size) - MIN_CHUNK_BITS;
+      // Add to local cache.
+#ifdef SNMALLOC_TRACING
+      message<1024>(
+        "DecayRange::dealloc_range: returning to local cache: {} on {}",
+        address_cast(base),
+        this);
+#endif
+      chunk_stack[slab_sizeclass][epoch].push(base);
+    }
+  };
+} // namespace snmalloc