Restructure the spinlock code:

- Add function-based spinlock API
- Improve comments
- Reduce bouncing with contested locks

include/xrpl/basics/CountedObject.h

@@ -1,29 +1,31 @@
 #pragma once
 
+#include <xrpl/beast/utility/instrumentation.h>
+
 #include <boost/core/type_name.hpp>
 
+#include <algorithm>
 #include <atomic>
+#include <cstddef>
+#include <cstdint>
+#include <iterator>
 #include <string>
+#include <type_traits>
 #include <utility>
-#include <vector>
 
 namespace xrpl {
 
 /**
  * Manages all counted object types.
+ *
+ * Counters register themselves on a lock-free intrusive list maintained
+ * by this object when constructed. Because counters are never destroyed
+ * or removed, the ABA problem does not apply.
+ *
+ * The registry is iterable as a forward range.
  */
 class CountedObjects
 {
-public:
-    static CountedObjects&
-    getInstance() noexcept;
-
-    using Entry = std::pair<std::string, int>;
-    using List = std::vector<Entry>;
-
-    [[nodiscard]] List
-    getCounts(int minimumThreshold) const;
-
 public:
     /**
      * Implementation for @ref CountedObject.
@@ -33,68 +35,147 @@ public:
     class Counter
     {
     public:
-        Counter(std::string name) noexcept : name_(std::move(name)), count_(0)
-        {
-            // Insert ourselves at the front of the lock-free linked list
-            CountedObjects& instance = CountedObjects::getInstance();
-            Counter* head = nullptr;
+        Counter(std::string name) noexcept;
 
-            do
-            {
-                head = instance.head_.load();
-                next_ = head;
-            } while (instance.head_.exchange(this) != head);
+        // Counters are intrusive list nodes whose addresses are published
+        // in the registry; they must never be copied or moved. The atomic
+        // members already force this, but we make it explicit.
+        Counter(Counter const&) = delete;
+        Counter&
+        operator=(Counter const&) = delete;
+        Counter(Counter&&) = delete;
+        Counter&
+        operator=(Counter&&) = delete;
 
-            ++instance.count_;
-        }
-
-        ~Counter() noexcept = default;
-
-        int
+        std::uint32_t
         increment() noexcept
         {
-            return ++count_;
+            auto const newCount = count_.fetch_add(1, std::memory_order::relaxed) + 1;
+            XRPL_ASSERT(newCount != 0, "xrpl::CountedObjects::Counter::increment : no overflow");
+
+            auto maxCount = maxCount_.load(std::memory_order::relaxed);
+
+            while (newCount > maxCount &&
+                   !maxCount_.compare_exchange_weak(maxCount, newCount, std::memory_order::relaxed))
+            {
+            }
+
+            return newCount;
         }
 
-        int
+        std::uint32_t
         decrement() noexcept
         {
-            return --count_;
+            auto const prev = count_.fetch_sub(1, std::memory_order::relaxed);
+            XRPL_ASSERT(prev != 0, "xrpl::CountedObjects::Counter::decrement : no underflow");
+            return prev - 1;
         }
 
-        [[nodiscard]] int
-        getCount() const noexcept
+        [[nodiscard]] std::uint32_t
+        count() const noexcept
         {
-            return count_.load();
+            return count_.load(std::memory_order::relaxed);
         }
 
-        [[nodiscard]] Counter*
-        getNext() const noexcept
+        [[nodiscard]] std::uint32_t
+        max() const noexcept
         {
-            return next_;
+            return std::max(
+                count_.load(std::memory_order::relaxed),
+                maxCount_.load(std::memory_order::relaxed));
         }
 
         [[nodiscard]] std::string const&
-        getName() const noexcept
+        name() const noexcept
         {
             return name_;
         }
 
     private:
+        friend class CountedObjects;
+
+        Counter* next_ = nullptr;
+        std::atomic<std::uint32_t> count_ = 0;
+        std::atomic<std::uint32_t> maxCount_ = 0;
         std::string const name_;
-        std::atomic<int> count_;
-        Counter* next_;
     };
 
-private:
-    CountedObjects() noexcept;
-    ~CountedObjects() noexcept = default;
+    class Iterator
+    {
+    public:
+        using value_type = Counter const;
+        using reference = value_type&;
+        using pointer = value_type*;
+        using difference_type = std::ptrdiff_t;
+        using iterator_category = std::forward_iterator_tag;
+
+        explicit Iterator(Counter* c = nullptr) noexcept : current_(c)
+        {
+        }
+
+        reference
+        operator*() const noexcept
+        {
+            return *current_;
+        }
+
+        pointer
+        operator->() const noexcept
+        {
+            return current_;
+        }
+
+        Iterator&
+        operator++() noexcept
+        {
+            current_ = current_->next_;
+            return *this;
+        }
+
+        Iterator
+        operator++(int) noexcept
+        {
+            auto tmp = *this;
+            ++*this;
+            return tmp;
+        }
+
+        bool
+        operator==(Iterator const&) const noexcept = default;
+
+    private:
+        Counter* current_;
+    };
+
+    constexpr CountedObjects() noexcept = default;
+
+    [[nodiscard]] auto
+    begin() const noexcept
+    {
+        return Iterator{head_.load(std::memory_order::acquire)};
+    }
+
+    [[nodiscard]] auto
+    end() const noexcept
+    {
+        return Iterator{};
+    }
 
 private:
-    std::atomic<int> count_;
-    std::atomic<Counter*> head_;
+    std::atomic<Counter*> head_ = nullptr;
 };
 
+/** The global counted object registry. */
+inline constinit CountedObjects gCountedObjects;
+
+inline CountedObjects::Counter::Counter(std::string name) noexcept
+    : next_(gCountedObjects.head_.load(std::memory_order::relaxed)), name_(std::move(name))
+{
+    while (!gCountedObjects.head_.compare_exchange_weak(
+        next_, this, std::memory_order::release, std::memory_order::relaxed))
+        ;
+}
+
 //------------------------------------------------------------------------------
 
 /**
@@ -103,27 +184,37 @@ private:
  * Derived classes have their instances counted automatically. This is used
  * for reporting purposes.
  *
+ * The constructors are private and `Object` is befriended so that the
+ * CRTP parameter must be the deriving class itself: a copy-paste error
+ * like `class B : public CountedObject<A>` fails to compile instead of
+ * silently polluting A's count.
+ *
+ * @note This class has no move operations by design: a derived class's
+ *       move constructor falls back to the copy constructor for this
+ *       base, so the newly created instance is counted. This keeps the
+ *       invariant that count is the number of outstanding subobjects.
+ *
+ * @warning Counted objects constructed during dynamic initialization of
+ *          other translation units may have their increments discarded when
+ *          counter itself is dynamically initialized. Do not create counted
+ *          objects before main() begins.
+ *
  * @ingroup basics
  */
 template <class Object>
+    requires std::is_class_v<Object>
 class CountedObject
 {
-private:
-    static auto&
-    getCounter() noexcept
-    {
-        static CountedObjects::Counter kC{boost::core::type_name<Object>()};
-        return kC;
-    }
+    static inline CountedObjects::Counter counter{boost::core::type_name<Object>()};
 
     CountedObject() noexcept
     {
-        getCounter().increment();
+        counter.increment();
     }
 
     CountedObject(CountedObject const&) noexcept
     {
-        getCounter().increment();
+        counter.increment();
     }
 
     CountedObject&
@@ -132,7 +223,7 @@ private:
 public:
     ~CountedObject() noexcept
     {
-        getCounter().decrement();
+        counter.decrement();
     }
 
     friend Object;

include/xrpl/basics/spinlock.h

@@ -4,88 +4,215 @@
 
 #include <xrpl/beast/utility/instrumentation.h>
 
+#include <boost/predef/architecture.h>
+
 #include <atomic>
+#include <concepts>
 #include <limits>
 #include <type_traits>
 
-#ifndef __aarch64__
+#if BOOST_ARCH_X86
 #include <immintrin.h>
 #endif
 
 namespace xrpl {
 
+/** An unsigned integral type suitable for use as a spinlock.
+
+    The type must be always lock-free when wrapped in std::atomic, so
+    that lock operations cannot themselves take a (library-level) lock.
+ */
+template <typename T>
+concept SpinlockValueType = std::is_unsigned_v<T> && std::atomic<T>::is_always_lock_free;
+
+/** A spinlock value type that additionally supports the atomic bitwise
+    operations required to pack multiple locks into a single integer.
+ */
+template <typename T>
+concept PackedSpinlockValueType = SpinlockValueType<T> && requires(std::atomic<T>& a, T v) {
+    { a.fetch_or(v) } -> std::same_as<T>;
+    { a.fetch_and(v) } -> std::same_as<T>;
+};
+
 namespace detail {
-/**
- * Inform the processor that we are in a tight spin-wait loop.
- *
- * Spinlocks caught in tight loops can result in the processor's pipeline
- * filling up with comparison operations, resulting in a misprediction at
- * the time the lock is finally acquired, necessitating pipeline flushing
- * which is ridiculously expensive and results in very high latency.
- *
- * This function instructs the processor to "pause" for some architecture
- * specific amount of time, to prevent this.
+
+/** Inform the processor that we are in a tight spin-wait loop.
+
+    Spinlocks caught in tight loops can result in the processor's pipeline
+    filling up with comparison operations, resulting in a misprediction at
+    the time the lock is finally acquired, necessitating pipeline flushing
+    which is ridiculously expensive and results in very high latency.
+
+    This function instructs the processor to "pause" for some architecture
+    specific amount of time, to prevent this.
  */
 inline void
 spinPause() noexcept
 {
-#ifdef __aarch64__
-    asm volatile("yield");
-#else
+#if BOOST_ARCH_X86
     _mm_pause();
+#elif BOOST_ARCH_ARM
+    asm volatile("yield" ::: "memory");
+#else
+#error No implementation available for spinPause to use
 #endif
 }
 
 }  // namespace detail
 
-/** @{ */
-/**
- * Classes to handle arrays of spinlocks packed into a single atomic integer:
- *
- * Packed spinlocks allow for tremendously space-efficient lock-sharding
- * but they come at a cost.
- *
- * First, the implementation is necessarily low-level and uses advanced
- * features like memory ordering and highly platform-specific tricks to
- * maximize performance. This imposes a significant and ongoing cost to
- * developers.
- *
- * Second, and perhaps most important, is that the packing of multiple
- * locks into a single integer which, albeit space-efficient, also has
- * performance implications stemming from data dependencies, increased
- * cache-coherency traffic between processors and heavier loads on the
- * processor's load/store units.
- *
- * To be sure, these locks can have advantages but they are definitely
- * not general purpose locks and should not be thought of or used that
- * way. The use cases for them are likely few and far between; without
- * a compelling reason to use them, backed by profiling data, it might
- * be best to use one of the standard locking primitives instead. Note
- * that in most common platforms, `std::mutex` is so heavily optimized
- * that it can, usually, outperform spinlocks.
- *
- * @tparam T An unsigned integral type (e.g. std::uint16_t)
- */
+//------------------------------------------------------------------------------
 
-/**
- * A class that grabs a single packed spinlock from an atomic integer.
- *
- * This class meets the requirements of Lockable:
- *     https://en.cppreference.com/w/cpp/named_req/Lockable
+/** Attempt to acquire a spinlock without blocking.
+
+    @note This interface is primarily intended for one-shot attempts to
+          acquire the lock. Avoid calling this function directly from a
+          loop and use @ref spinLock instead.
+
+    @tparam T An unsigned integral type.
+    @param lock The atomic variable used as the lock.
+    @return true if the lock was acquired, false if it was already held.
  */
-template <class T>
+template <SpinlockValueType T>
+[[nodiscard]] bool
+spinTryLock(std::atomic<T>& lock) noexcept
+{
+    // A compare-exchange is required here, not an unconditional exchange:
+    // a failed attempt must not modify the lock word, in case the atomic
+    // is shared with PackedSpinlock).
+    T expected = 0;
+
+    return lock.compare_exchange_strong(
+        expected,
+        std::numeric_limits<T>::max(),
+        std::memory_order::acquire,
+        std::memory_order::relaxed);
+    ;
+}
+
+/** Acquire a spinlock, blocking until available.
+
+    Uses a TTAS (test-and-test-and-set) pattern, so waiters share the cache
+    line read-only, helping to avoid unnecessary coherency traffic.
+
+    @tparam T An unsigned integral type.
+    @param lock The atomic variable used as the lock.
+ */
+template <SpinlockValueType T>
+void
+spinLock(std::atomic<T>& lock) noexcept
+{
+    do
+    {
+        // Relaxed ordering is sufficient for the spin: this load is only
+        // a filter. The acquire on the successful exchange in spinTryLock
+        // is what synchronizes the critical section.
+        while (lock.load(std::memory_order::relaxed) != 0)
+            detail::spinPause();
+    } while (!spinTryLock(lock));
+}
+
+/** Release a spinlock.
+
+    @tparam T An unsigned integral type.
+    @param lock The atomic variable used as the lock.
+ */
+template <SpinlockValueType T>
+void
+spinUnlock(std::atomic<T>& lock) noexcept
+{
+    lock.store(0, std::memory_order::release);
+}
+
+//------------------------------------------------------------------------------
+
+/** A Lockable interface to a spinlock implemented on top of an atomic.
+
+    @tparam T An unsigned integral type.
+
+    @note Using `PackedSpinlock` and `Spinlock` against the same underlying
+          atomic integer is possible but can result in `Spinlock` not being
+          able to acquire the lock during periods of high contention due to
+          the way the two locks operate: `Spinlock` spins and tries to grab
+          all the bits at once, whereas any given `PackedSpinlock` instance
+          only tries to grab one bit at a time. Caveat emptor.
+
+    This class meets the requirements of Lockable:
+        https://en.cppreference.com/w/cpp/named_req/Lockable
+ */
+template <SpinlockValueType T>
+class Spinlock
+{
+    std::atomic<T>& lock_;
+
+public:
+    Spinlock(Spinlock const&) = delete;
+    Spinlock&
+    operator=(Spinlock const&) = delete;
+
+    /** Construct a spinlock handle.
+
+        @param lock The atomic integer to spin against.
+
+        @note For performance reasons, you should strive to have `lock` be
+              on a cacheline by itself.
+     */
+    explicit Spinlock(std::atomic<T>& lock) noexcept : lock_(lock)
+    {
+    }
+
+    [[nodiscard]] bool
+    try_lock() noexcept  // NOLINT(readability-identifier-naming)
+    {
+        return spinTryLock(lock_);
+    }
+
+    void
+    lock() noexcept
+    {
+        spinLock(lock_);
+    }
+
+    void
+    unlock() noexcept
+    {
+        spinUnlock(lock_);
+    }
+};
+
+//------------------------------------------------------------------------------
+
+/** A Lockable interface to a packed spinlock implemented on top of an atomic.
+
+    Packed spinlocks offer tremendous space-efficient lock-sharding but
+    they come at a cost.
+
+    First, the implementation is necessarily low-level and uses advanced
+    features like memory ordering and highly platform-specific tricks to
+    maximize performance. This imposes a significant and ongoing cost to
+    developers.
+
+    Second, and perhaps most important, is that the packing of multiple
+    locks into a single integer which, albeit space-efficient, also has
+    performance implications stemming from data dependencies, increased
+    cache-coherency traffic between processors and heavier loads on the
+    processor's load/store units.
+
+    To be sure, these locks can have advantages but they are definitely
+    not general purpose locks and should not be thought of or used that
+    way. The use cases for them are likely few and far between; without
+    a compelling reason to use them, backed by profiling data, it might
+    be best to use one of the standard locking primitives instead. Note
+    that in most common platforms, `std::mutex` is so heavily optimized
+    that it can, usually, outperform spinlocks.
+
+    @tparam T An unsigned integral type (e.g. std::uint16_t)
+
+    This class meets the requirements of Lockable:
+        https://en.cppreference.com/w/cpp/named_req/Lockable
+ */
+template <PackedSpinlockValueType T>
 class PackedSpinlock
 {
-    // clang-format off
-    static_assert(std::is_unsigned_v<T>);
-    static_assert(std::atomic<T>::is_always_lock_free);
-    static_assert(
-        std::is_same_v<decltype(std::declval<std::atomic<T>&>().fetch_or(0)), T> &&
-        std::is_same_v<decltype(std::declval<std::atomic<T>&>().fetch_and(0)), T>,
-        "std::atomic<T>::fetch_and(T) and std::atomic<T>::fetch_and(T) are required by packed_spinlock");
-    // clang-format on
-
-private:
     std::atomic<T>& bits_;
     T const mask_;
 
@@ -94,120 +221,49 @@ public:
     PackedSpinlock&
     operator=(PackedSpinlock const&) = delete;
 
-    /**
-     * A single spinlock packed inside the specified atomic
-     *
-     * @param lock The atomic integer inside which the spinlock is packed.
-     * @param index The index of the spinlock this object acquires.
-     *
-     * @note For performance reasons, you should strive to have `lock` be
-     *       on a cacheline by itself.
+    /** Construct a packed spinlock handle for a single bit.
+
+        @param lock The atomic integer inside which the spinlock is packed.
+        @param index The index of the spinlock this object acquires.
+
+        @note For performance reasons, you should strive to have `lock` be
+              on a cacheline by itself.
      */
-    PackedSpinlock(std::atomic<T>& lock, int index) : bits_(lock), mask_(static_cast<T>(1) << index)
-    {
-        XRPL_ASSERT(
-            index >= 0 && (mask_ != 0),
-            "xrpl::PackedSpinlock::PackedSpinlock : valid index and mask");
-    }
-
-    [[nodiscard]] bool
-    try_lock()  // NOLINT(readability-identifier-naming)
-    {
-        return (bits_.fetch_or(mask_, std::memory_order_acquire) & mask_) == 0;
-    }
-
-    void
-    lock()
-    {
-        while (!try_lock())
-        {
-            // The use of relaxed memory ordering here is intentional and
-            // serves to help reduce cache coherency traffic during times
-            // of contention by avoiding writes that would definitely not
-            // result in the lock being acquired.
-            while ((bits_.load(std::memory_order_relaxed) & mask_) != 0)
-                detail::spinPause();
-        }
-    }
-
-    void
-    unlock()
-    {
-        bits_.fetch_and(~mask_, std::memory_order_release);
-    }
-};
-
-/**
- * A spinlock implemented on top of an atomic integer.
- *
- * @note Using `packed_spinlock` and `spinlock` against the same underlying
- *       atomic integer can result in `spinlock` not being able to actually
- *       acquire the lock during periods of high contention, because of how
- *       the two locks operate: `spinlock` will spin trying to grab all the
- *       bits at once, whereas any given `packed_spinlock` will only try to
- *       grab one bit at a time. Caveat emptor.
- *
- * This class meets the requirements of Lockable:
- *     https://en.cppreference.com/w/cpp/named_req/Lockable
- */
-template <class T>
-class Spinlock
-{
-    static_assert(std::is_unsigned_v<T>);
-    static_assert(std::atomic<T>::is_always_lock_free);
-
-private:
-    std::atomic<T>& lock_;
-
-public:
-    Spinlock(Spinlock const&) = delete;
-    Spinlock&
-    operator=(Spinlock const&) = delete;
-
-    /**
-     * Grabs the
-     *
-     * @param lock The atomic integer to spin against.
-     *
-     * @note For performance reasons, you should strive to have `lock` be
-     *       on a cacheline by itself.
-     */
-    Spinlock(std::atomic<T>& lock) : lock_(lock)
+    PackedSpinlock(std::atomic<T>& lock, int index) noexcept
+        : bits_(lock), mask_([index]() {
+            XRPL_ASSERT(
+                index >= 0 && index < std::numeric_limits<T>::digits,
+                "xrpl::PackedSpinlock::PackedSpinlock : valid index");
+            return static_cast<T>(T{1} << index);
+        }())
     {
     }
 
     [[nodiscard]] bool
-    try_lock()  // NOLINT(readability-identifier-naming)
+    try_lock() noexcept  // NOLINT(readability-identifier-naming)
     {
-        T expected = 0;
-
-        return lock_.compare_exchange_weak(
-            expected,
-            std::numeric_limits<T>::max(),
-            std::memory_order_acquire,
-            std::memory_order_relaxed);
+        return (bits_.fetch_or(mask_, std::memory_order::acquire) & mask_) == 0;
     }
 
     void
-    lock()
+    lock() noexcept
     {
-        while (!try_lock())
+        do
         {
             // The use of relaxed memory ordering here is intentional and
             // serves to help reduce cache coherency traffic during times
-            // of contention by avoiding writes that would definitely not
-            // result in the lock being acquired.
-            while (lock_.load(std::memory_order_relaxed) != 0)
+            // of contention by avoiding writes that are unlikely to grab
+            // the requested lock.
+            while ((bits_.load(std::memory_order::relaxed) & mask_) != 0)
                 detail::spinPause();
-        }
+        } while (!try_lock());
     }
 
     void
-    unlock()
+    unlock() noexcept
     {
-        lock_.store(0, std::memory_order_release);
+        bits_.fetch_and(~mask_, std::memory_order::release);
     }
 };
-/** @} */
 
 }  // namespace xrpl

include/xrpl/protocol/AccountID.h

@@ -101,22 +101,6 @@ operator<<(std::ostream& os, AccountID const& x)
     return os;
 }
 
-/**
- * Initialize the global cache used to map AccountID to base58 conversions.
- *
- * The cache is optional and need not be initialized. But because conversion
- * is expensive (it requires a SHA-256 operation) in most cases the overhead
- * of the cache is worth the benefit.
- *
- * @param count The number of entries the cache should accommodate. Zero will
- *              disable the cache, releasing any memory associated with it.
- *
- * @note The function will only initialize the cache the first time it is
- *       invoked. Subsequent invocations do nothing.
- */
-void
-initAccountIdCache(std::size_t count);
-
 }  // namespace xrpl
 
 //------------------------------------------------------------------------------

include/xrpl/protocol/jss.h

@@ -392,6 +392,7 @@ JSS(max_ledger);                  // in/out: LedgerCleaner
 JSS(max_queue_size);              // out: TxQ
 JSS(max_spend_drops);             // out: AccountInfo
 JSS(max_spend_drops_total);       // out: AccountInfo
+JSS(maximum);
 JSS(mean);                        // out: get_aggregate_price
 JSS(median);                      // out: get_aggregate_price
 JSS(median_fee);                  // out: TxQ

src/libxrpl/basics/CountedObject.cpp

@@ -1,39 +0,0 @@
-#include <xrpl/basics/CountedObject.h>
-
-#include <algorithm>
-
-namespace xrpl {
-
-CountedObjects&
-CountedObjects::getInstance() noexcept
-{
-    static CountedObjects kInstance;
-
-    return kInstance;
-}
-
-CountedObjects::CountedObjects() noexcept : count_(0), head_(nullptr)
-{
-}
-
-CountedObjects::List
-CountedObjects::getCounts(int minimumThreshold) const
-{
-    List counts;
-
-    // When other operations are concurrent, the count
-    // might be temporarily less than the actual count.
-    counts.reserve(count_.load());
-
-    for (auto* ctr = head_.load(); ctr != nullptr; ctr = ctr->getNext())
-    {
-        if (ctr->getCount() >= minimumThreshold)
-            counts.emplace_back(ctr->getName(), ctr->getCount());
-    }
-
-    std::ranges::sort(counts);
-
-    return counts;
-}
-
-}  // namespace xrpl

src/libxrpl/protocol/AccountID.cpp

@@ -8,6 +8,8 @@
 #include <xrpl/protocol/digest.h>
 #include <xrpl/protocol/tokens.h>
 
+#include <algorithm>
+#include <array>
 #include <atomic>
 #include <cstdint>
 #include <cstring>
@@ -19,85 +21,136 @@
 
 namespace xrpl {
 
-namespace detail {
+namespace {
 
-/**
- * Caches the base58 representations of AccountIDs
+/** If true, we implement a cache for account ID lookups.
+
+    This is generally a good idea and can improve performance, especially
+    when serving client requests.
+
+    The cache may generate warnings under some sanitizers; if you want to
+    disable caching, set this variable to false.
  */
-class AccountIdCache
+constexpr bool cacheAccountIDs = true;
+
+/** Represents a cached entry for the given AccountID.
+
+    We keep the size and alignment of the structure at 64 bytes to
+    make sure it fits within a cache line and avoid false sharing.
+ */
+struct alignas(64) CachedAccountID
 {
-private:
-    struct CachedAccountID
-    {
-        AccountID id;
-        char encoding[40] = {0};
-    };
+    /** The sequence lock for the entry.
 
-    // The actual cache
-    std::vector<CachedAccountID> cache_;
+        An even value means that the data is stable, whereas an odd
+        value indicates the data is being actively modified.
 
-    // We use a hash function designed to resist algorithmic complexity attacks
-    HardenedHash<> hasher_;
+        We use a 32-bit value to ensure that the counter will never
+        return to an observed value during a read window.
+     */
+    std::atomic<std::uint32_t> seq{0};
 
-    // 64 spinlocks, packed into a single 64-bit value
-    std::atomic<std::uint64_t> locks_ = 0;
+    /** The account id. */
+    AccountID id{};
 
-public:
-    AccountIdCache(std::size_t count) : cache_(count)
-    {
-        // This is non-binding, but we try to avoid wasting memory that
-        // is caused by overallocation.
-        cache_.shrink_to_fit();
-    }
+    /** The length of the cached encoding.
 
-    std::string
-    toBase58(AccountID const& id)
-    {
-        auto const index = hasher_(id) % cache_.size();
+        This field disambiguates collisions between ACCOUNT_ZERO and
+        never-written-to slots, since they would otherwise share the
+        same encoded bits.
+     */
+    std::uint8_t len = 0;
 
-        PackedSpinlock sl(locks_, index % 64);
-
-        {
-            std::scoped_lock const lock(sl);
-
-            // The check against the first character of the encoding ensures
-            // that we don't mishandle the case of the all-zero account:
-            if (cache_[index].encoding[0] != 0 && cache_[index].id == id)
-                return cache_[index].encoding;
-        }
-
-        auto ret = encodeBase58Token(TokenType::AccountID, id.data(), id.size());
-
-        XRPL_ASSERT(ret.size() <= 38, "xrpl::detail::AccountIdCache : maximum result size");
-
-        {
-            std::scoped_lock const lock(sl);
-            cache_[index].id = id;
-            std::strcpy(cache_[index].encoding, ret.c_str());
-        }
-
-        return ret;
-    }
+    /** The actual string encoding. */
+    char encoding[39] = {0};
 };
 
-}  // namespace detail
+static_assert(sizeof(CachedAccountID) == 64 && alignof(CachedAccountID) == 64);
 
-static std::unique_ptr<detail::AccountIdCache> gAccountIdCache;
+// The actual cache to use: a simple, direct-mapped, best-effort construct.
+//
+// We do not use a hardened hash for this because an AccountID is already
+// just cryptographically generated (i.e. uniformly distributed) bits and
+// there is no amplification, since collisions do not carry an additional
+// cost.
+//
+// The size is a "nice" power of 2 because it makes the calculation of an
+// index easier.
+constinit std::array<CachedAccountID, 65536> gAccountIdCache{};
 
-void
-initAccountIdCache(std::size_t count)
-{
-    if (!gAccountIdCache && count != 0)
-        gAccountIdCache = std::make_unique<detail::AccountIdCache>(count);
-}
+}  // namespace
 
 std::string
 toBase58(AccountID const& v)
 {
-    if (gAccountIdCache)
-        return gAccountIdCache->toBase58(v);
+    // Since the input bits are generated by a cryptographically secure
+    // hash construct, we can just pick any subset of them, no muss, no
+    // fuss.
+    auto const index = [&v]() -> std::uint16_t {
+        auto const data = v.data();
 
-    return encodeBase58Token(TokenType::AccountID, v.data(), v.size());
+        return static_cast<std::uint16_t>(data[0]) | (static_cast<std::uint16_t>(data[1]) << 8);
+    }();
+
+    if constexpr (cacheAccountIDs)
+    {
+        auto& slot = gAccountIdCache[index];
+
+        // This is the optimistic, store-free read path: a collision with a
+        // concurrent writer will be detected because the seq recheck fails
+        // and we fall through to the slow path.
+        //
+        // These unsynchronized reads of the slot fields are, formally, data
+        // races and, technically speaking, result in undefined behavior and
+        // a tool like TSAN might flag them. These should not have an impact
+        // on the correctness of the returned data: the sequence lock checks
+        // ensure we discard any data that was in flux.
+        if (auto const s0 = slot.seq.load(std::memory_order::acquire);
+            (s0 & 1) == 0 && slot.len != 0 && slot.id == v)
+        {
+            char buf[38];
+            auto const len = std::min<std::size_t>(slot.len, sizeof(buf));
+            std::memcpy(buf, slot.encoding, len);
+
+            // Prevents the loads above from sinking below the recheck; the
+            // acquire on s0 constrains later operations, not earlier ones.
+            std::atomic_thread_fence(std::memory_order::acquire);
+
+            // Everything read from the slot — the gates above and the bytes in
+            // buf — precedes this recheck and is validated by it. No slot access
+            // may follow it.
+            if (slot.seq.load(std::memory_order::relaxed) == s0)
+                return std::string(buf, len);
+        }
+    }
+
+    auto ret = encodeBase58Token(TokenType::AccountID, v.data(), v.size());
+
+    XRPL_ASSERT(!ret.empty() && ret.size() <= 38, "xrpl::toBase58 : maximum result size");
+
+    if constexpr (cacheAccountIDs)
+    {
+        auto& slot = gAccountIdCache[index];
+
+        // The CAS to an odd value excludes other writers; if it fails, another
+        // thread wrote or is writing into this slot. That is not a problem: we
+        // just don't store our result into the cache this time.
+        if (auto s0 = slot.seq.load(std::memory_order::relaxed); (s0 & 1) == 0 &&
+            slot.seq.compare_exchange_strong(s0, s0 + 1, std::memory_order::acquire))
+        {
+            slot.id = v;
+            slot.len = static_cast<std::uint8_t>(ret.size());
+            std::copy_n(ret.data(), ret.size(), slot.encoding);
+
+            // The acquire on the CAS above keeps the writes from becoming visible
+            // before the seqlock transitions to an odd value, which means readers
+            // can detect an in-progress write. This release publishes them before
+            // the seqlock transitions to an even value.
+            slot.seq.store(s0 + 2, std::memory_order::release);
+        }
+    }
+
+    return ret;
 }
 
 template <>

src/test/app/Regression_test.cpp

@@ -305,12 +305,11 @@ struct Regression_test : public beast::unit_test::Suite
                 return digest.asUInt256();
             }();
 
-            auto const mapCounts = [&](CountedObjects::List const& list) {
-                std::map<std::string, int> result;
-                for (auto const& e : list)
-                {
-                    result[e.first] = e.second;
-                }
+            auto const mapCounts = [] {
+                std::map<std::string, std::uint32_t> result;
+
+                for (auto const& c : gCountedObjects)
+                    result.emplace(c.name(), c.count());
 
                 return result;
             };
@@ -319,11 +318,11 @@ struct Regression_test : public beast::unit_test::Suite
             {
                 auto& cache = env.app().getCachedSLEs();
                 cache.del(*digest, false);  // NOLINT(bugprone-unchecked-optional-access)
-                auto const beforeCounts = mapCounts(CountedObjects::getInstance().getCounts(0));
+                auto const beforeCounts = mapCounts();
 
                 env(check::cash(alice, bobIndex, check::DeliverMin(XRP(100))), Ter(tecNO_ENTRY));
 
-                auto const afterCounts = mapCounts(CountedObjects::getInstance().getCounts(0));
+                auto const afterCounts = mapCounts();
 
                 using namespace std::string_literals;
                 BEAST_EXPECT(

src/test/rpc/GetCounts_test.cpp

@@ -9,6 +9,7 @@
 #include <xrpl/json/json_value.h>
 #include <xrpl/protocol/jss.h>
 
+#include <algorithm>
 #include <thread>
 
 namespace xrpl {
@@ -52,36 +53,59 @@ class GetCounts_test : public beast::unit_test::Suite
             // check counts, default params
             result = env.rpc("get_counts")[jss::result];
             BEAST_EXPECT(result[jss::status] == "success");
-            // compare with values reported by CountedObjects
-            auto const& objectCounts = CountedObjects::getInstance().getCounts(10);
-            for (auto const& it : objectCounts)
+
+            // compare with values reported by the registry
+            auto const& counters = result[jss::counters];
+
+            for (auto const& c : gCountedObjects)
             {
-                BEAST_EXPECTS(result.isMember(it.first), it.first);
-                BEAST_EXPECTS(result[it.first].asInt() == it.second, it.first);
+                if (auto const count = c.count(); count >= 10)
+                {
+                    BEAST_EXPECTS(counters.isMember(c.name()), c.name());
+                    BEAST_EXPECTS(counters[c.name()][jss::current].asUInt() == count, c.name());
+                }
             }
+
             BEAST_EXPECT(!result.isMember(jss::local_txs));
         }
-
         {
             // make request with min threshold 100 and verify
-            // that only STObject and NodeObject are reported
+            // that only counters at or above the threshold are reported
             result = env.rpc("get_counts", "100")[jss::result];
             BEAST_EXPECT(result[jss::status] == "success");
 
-            // compare with values reported by CountedObjects
-            auto const& objectCounts = CountedObjects::getInstance().getCounts(100);
-            for (auto const& it : objectCounts)
-            {
-                BEAST_EXPECTS(result.isMember(it.first), it.first);
-                BEAST_EXPECTS(result[it.first].asInt() == it.second, it.first);
-            }
-            BEAST_EXPECT(!result.isMember("Transaction"));
-            BEAST_EXPECT(!result.isMember("STTx"));
-            BEAST_EXPECT(!result.isMember("STArray"));
-            BEAST_EXPECT(!result.isMember("HashRouterEntry"));
-            BEAST_EXPECT(!result.isMember("STLedgerEntry"));
-        }
+            auto const& counters = result[jss::counters];
 
+            // every registry counter at/above threshold must be reported, with
+            // a matching current value
+            for (auto const& c : gCountedObjects)
+            {
+                if (auto const count = c.count(); count >= 100)
+                {
+                    BEAST_EXPECTS(counters.isMember(c.name()), c.name());
+                    BEAST_EXPECTS(counters[c.name()][jss::current].asUInt() == count, c.name());
+                }
+            }
+
+            // conversely, every reported entry must be a known counter that met
+            // the threshold, and its maximum must be consistent
+            for (auto const& name : counters.getMemberNames())
+            {
+                auto const it = std::ranges::find_if(
+                    gCountedObjects, [&](auto const& c) { return c.name() == name; });
+                BEAST_EXPECTS(it != gCountedObjects.end(), name);
+
+                auto const& entry = counters[name];
+                BEAST_EXPECTS(entry[jss::current].asUInt() >= 100, name);
+                BEAST_EXPECTS(entry[jss::maximum].asUInt() >= entry[jss::current].asUInt(), name);
+            }
+
+            BEAST_EXPECT(!counters.isMember("xrpl::Transaction"));
+            BEAST_EXPECT(!counters.isMember("xrpl::STTx"));
+            BEAST_EXPECT(!counters.isMember("xrpl::STArray"));
+            BEAST_EXPECT(!counters.isMember("xrpl::HashRouterEntry"));
+            BEAST_EXPECT(!counters.isMember("xrpl::STLedgerEntry"));
+        }
         {
             // local_txs field will exist when there are open Txs
             env(pay(alice, bob, alice["USD"](5)));

src/xrpld/app/main/Application.cpp

@@ -473,8 +473,6 @@ public:
               getIoContext())
         , grpcServer_(std::make_unique<GRPCServer>(*this))
     {
-        initAccountIdCache(config_->getValueFor(SizedItem::AccountIdCacheSize));
-
         add(resourceManager_.get());
 
         //

src/xrpld/core/Config.h

@@ -41,7 +41,6 @@ enum class SizedItem : std::size_t {
     OpenFinalLimit,
     BurstSize,
     RamSizeGb,
-    AccountIdCacheSize,
 };
 
 /**

src/xrpld/core/detail/Config.cpp

@@ -111,7 +111,7 @@ namespace xrpl {
 
 // clang-format off
 // The configurable node sizes are "tiny", "small", "medium", "large", "huge"
-inline constexpr std::array<std::pair<SizedItem, std::array<int, 5>>, 13>
+inline constexpr std::array<std::pair<SizedItem, std::array<int, 5>>, 12>
 kSizedItems
 {{
     // FIXME: We should document each of these items, explaining exactly
@@ -130,8 +130,7 @@ kSizedItems
     {SizedItem::LgrDbCache,         {{      4,       8,      16,      32,     128 }}},
     {SizedItem::OpenFinalLimit,     {{      8,      16,      32,      64,     128 }}},
     {SizedItem::BurstSize,          {{      4,       8,      16,      32,      48 }}},
-    {SizedItem::RamSizeGb,          {{      6,       8,      12,      24,       0 }}},
-    {SizedItem::AccountIdCacheSize, {{  20047,   50053,   77081,  150061,  300007 }}}
+    {SizedItem::RamSizeGb,          {{      6,       8,      12,      24,       0 }}}
 }};
 // clang-format on
 

src/xrpld/rpc/handlers/admin/status/GetCounts.cpp

@@ -4,6 +4,7 @@
 #include <xrpld/app/rdb/backend/SQLiteDatabase.h>
 #include <xrpld/rpc/Context.h>
 
+#include <xrpl/basics/CountedObject.h>
 #include <xrpl/basics/UptimeClock.h>
 #include <xrpl/json/json_forwards.h>
 #include <xrpl/json/json_value.h>
@@ -11,9 +12,11 @@
 #include <xrpl/protocol/jss.h>
 #include <xrpl/server/NetworkOPs.h>
 
+#include <algorithm>
 #include <chrono>
 #include <cstddef>
 #include <string>
+#include <utility>
 
 namespace xrpl {
 
@@ -43,16 +46,26 @@ textTime(
 }
 
 json::Value
-getCountsJson(Application& app, int minObjectCount)
+getCountsJson(Application& app, std::uint32_t minObjectCount)
 {
-    auto objectCounts = CountedObjects::getInstance().getCounts(minObjectCount);
+    json::Value ret = json::ValueType::Object;
 
-    json::Value ret(json::ValueType::Object);
+    ret[jss::counters] = [minObjectCount] {
+        json::Value ctrs = json::ValueType::Object;
 
-    for (auto const& [k, v] : objectCounts)
-    {
-        ret[k] = v;
-    }
+        for (auto const& c : gCountedObjects)
+        {
+            if (auto const count = c.count(); count >= minObjectCount)
+            {
+                json::Value obj(json::ValueType::Object);
+                obj[jss::current] = count;
+                obj[jss::maximum] = std::max(count, c.max());
+                ctrs[c.name()] = std::move(obj);
+            }
+        }
+
+        return ctrs;
+    }();
 
     if (app.config().useTxTables())
     {
@@ -113,7 +126,7 @@ getCountsJson(Application& app, int minObjectCount)
 json::Value
 doGetCounts(rpc::JsonContext& context)
 {
-    int minCount = 10;
+    std::uint32_t minCount = 10;
 
     if (context.params.isMember(jss::min_count))
         minCount = context.params[jss::min_count].asUInt();

src/xrpld/rpc/handlers/admin/status/GetCounts.h

@@ -4,9 +4,11 @@
 
 #include <xrpl/json/json_value.h>
 
+#include <cstdint>
+
 namespace xrpl {
 
 json::Value
-getCountsJson(Application& app, int minObjectCount);
+getCountsJson(Application& app, std::uint32_t minObjectCount);
 
 }  // namespace xrpl