[core] Add memory pool to scheduler to reduce heap fragmentation (#10536)

esphome/core/scheduler.cpp

@@ -14,7 +14,20 @@ namespace esphome {
 
 static const char *const TAG = "scheduler";
 
-static const uint32_t MAX_LOGICALLY_DELETED_ITEMS = 10;
+// Memory pool configuration constants
+// Pool size of 5 matches typical usage patterns (2-4 active timers)
+// - Minimal memory overhead (~250 bytes on ESP32)
+// - Sufficient for most configs with a couple sensors/components
+// - Still prevents heap fragmentation and allocation stalls
+// - Complex setups with many timers will just allocate beyond the pool
+// See https://github.com/esphome/backlog/issues/52
+static constexpr size_t MAX_POOL_SIZE = 5;
+
+// Maximum number of logically deleted (cancelled) items before forcing cleanup.
+// Set to 5 to match the pool size - when we have as many cancelled items as our
+// pool can hold, it's time to clean up and recycle them.
+static constexpr uint32_t MAX_LOGICALLY_DELETED_ITEMS = 5;
+
 // Half the 32-bit range - used to detect rollovers vs normal time progression
 static constexpr uint32_t HALF_MAX_UINT32 = std::numeric_limits<uint32_t>::max() / 2;
 // max delay to start an interval sequence
@@ -79,8 +92,28 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
     return;
   }
 
+  // Get fresh timestamp BEFORE taking lock - millis_64_ may need to acquire lock itself
+  const uint64_t now = this->millis_64_(millis());
+
+  // Take lock early to protect scheduler_item_pool_ access
+  LockGuard guard{this->lock_};
+
   // Create and populate the scheduler item
-  auto item = make_unique<SchedulerItem>();
+  std::unique_ptr<SchedulerItem> item;
+  if (!this->scheduler_item_pool_.empty()) {
+    // Reuse from pool
+    item = std::move(this->scheduler_item_pool_.back());
+    this->scheduler_item_pool_.pop_back();
+#ifdef ESPHOME_DEBUG_SCHEDULER
+    ESP_LOGD(TAG, "Reused item from pool (pool size now: %zu)", this->scheduler_item_pool_.size());
+#endif
+  } else {
+    // Allocate new if pool is empty
+    item = make_unique<SchedulerItem>();
+#ifdef ESPHOME_DEBUG_SCHEDULER
+    ESP_LOGD(TAG, "Allocated new item (pool empty)");
+#endif
+  }
   item->component = component;
   item->set_name(name_cstr, !is_static_string);
   item->type = type;
@@ -99,7 +132,6 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
   // Single-core platforms don't need thread-safe defer handling
   if (delay == 0 && type == SchedulerItem::TIMEOUT) {
     // Put in defer queue for guaranteed FIFO execution
-    LockGuard guard{this->lock_};
     if (!skip_cancel) {
       this->cancel_item_locked_(component, name_cstr, type);
     }
@@ -108,9 +140,6 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
   }
 #endif /* not ESPHOME_THREAD_SINGLE */
 
-  // Get fresh timestamp for new timer/interval - ensures accurate scheduling
-  const auto now = this->millis_64_(millis());  // Fresh millis() call
-
   // Type-specific setup
   if (type == SchedulerItem::INTERVAL) {
     item->interval = delay;
@@ -142,8 +171,6 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
   }
 #endif /* ESPHOME_DEBUG_SCHEDULER */
 
-  LockGuard guard{this->lock_};
-
   // For retries, check if there's a cancelled timeout first
   if (is_retry && name_cstr != nullptr && type == SchedulerItem::TIMEOUT &&
       (has_cancelled_timeout_in_container_(this->items_, component, name_cstr, /* match_retry= */ true) ||
@@ -319,6 +346,8 @@ void HOT Scheduler::call(uint32_t now) {
     if (!this->should_skip_item_(item.get())) {
       this->execute_item_(item.get(), now);
     }
+    // Recycle the defer item after execution
+    this->recycle_item_(std::move(item));
   }
 #endif /* not ESPHOME_THREAD_SINGLE */
 
@@ -338,11 +367,11 @@ void HOT Scheduler::call(uint32_t now) {
 #ifdef ESPHOME_THREAD_MULTI_ATOMICS
     const auto last_dbg = this->last_millis_.load(std::memory_order_relaxed);
     const auto major_dbg = this->millis_major_.load(std::memory_order_relaxed);
-    ESP_LOGD(TAG, "Items: count=%zu, now=%" PRIu64 " (%" PRIu16 ", %" PRIu32 ")", this->items_.size(), now_64,
-             major_dbg, last_dbg);
+    ESP_LOGD(TAG, "Items: count=%zu, pool=%zu, now=%" PRIu64 " (%" PRIu16 ", %" PRIu32 ")", this->items_.size(),
+             this->scheduler_item_pool_.size(), now_64, major_dbg, last_dbg);
 #else  /* not ESPHOME_THREAD_MULTI_ATOMICS */
-    ESP_LOGD(TAG, "Items: count=%zu, now=%" PRIu64 " (%" PRIu16 ", %" PRIu32 ")", this->items_.size(), now_64,
-             this->millis_major_, this->last_millis_);
+    ESP_LOGD(TAG, "Items: count=%zu, pool=%zu, now=%" PRIu64 " (%" PRIu16 ", %" PRIu32 ")", this->items_.size(),
+             this->scheduler_item_pool_.size(), now_64, this->millis_major_, this->last_millis_);
 #endif /* else ESPHOME_THREAD_MULTI_ATOMICS */
     // Cleanup before debug output
     this->cleanup_();
@@ -355,9 +384,10 @@ void HOT Scheduler::call(uint32_t now) {
       }
 
       const char *name = item->get_name();
-      ESP_LOGD(TAG, "  %s '%s/%s' interval=%" PRIu32 " next_execution in %" PRIu64 "ms at %" PRIu64,
+      bool is_cancelled = is_item_removed_(item.get());
+      ESP_LOGD(TAG, "  %s '%s/%s' interval=%" PRIu32 " next_execution in %" PRIu64 "ms at %" PRIu64 "%s",
                item->get_type_str(), item->get_source(), name ? name : "(null)", item->interval,
-               item->next_execution_ - now_64, item->next_execution_);
+               item->next_execution_ - now_64, item->next_execution_, is_cancelled ? " [CANCELLED]" : "");
 
       old_items.push_back(std::move(item));
     }
@@ -372,8 +402,13 @@ void HOT Scheduler::call(uint32_t now) {
   }
 #endif /* ESPHOME_DEBUG_SCHEDULER */
 
-  // If we have too many items to remove
-  if (this->to_remove_ > MAX_LOGICALLY_DELETED_ITEMS) {
+  // Cleanup removed items before processing
+  // First try to clean items from the top of the heap (fast path)
+  this->cleanup_();
+
+  // If we still have too many cancelled items, do a full cleanup
+  // This only happens if cancelled items are stuck in the middle/bottom of the heap
+  if (this->to_remove_ >= MAX_LOGICALLY_DELETED_ITEMS) {
     // We hold the lock for the entire cleanup operation because:
     // 1. We're rebuilding the entire items_ list, so we need exclusive access throughout
     // 2. Other threads must see either the old state or the new state, not intermediate states
@@ -383,10 +418,13 @@ void HOT Scheduler::call(uint32_t now) {
 
     std::vector<std::unique_ptr<SchedulerItem>> valid_items;
 
-    // Move all non-removed items to valid_items
+    // Move all non-removed items to valid_items, recycle removed ones
     for (auto &item : this->items_) {
-      if (!item->remove) {
+      if (!is_item_removed_(item.get())) {
         valid_items.push_back(std::move(item));
+      } else {
+        // Recycle removed items
+        this->recycle_item_(std::move(item));
       }
     }
 
@@ -396,9 +434,6 @@ void HOT Scheduler::call(uint32_t now) {
     std::make_heap(this->items_.begin(), this->items_.end(), SchedulerItem::cmp);
     this->to_remove_ = 0;
   }
-
-  // Cleanup removed items before processing
-  this->cleanup_();
   while (!this->items_.empty()) {
     // use scoping to indicate visibility of `item` variable
     {
@@ -472,6 +507,9 @@ void HOT Scheduler::call(uint32_t now) {
         // Add new item directly to to_add_
         // since we have the lock held
         this->to_add_.push_back(std::move(item));
+      } else {
+        // Timeout completed - recycle it
+        this->recycle_item_(std::move(item));
       }
 
       has_added_items |= !this->to_add_.empty();
@@ -485,7 +523,9 @@ void HOT Scheduler::call(uint32_t now) {
 void HOT Scheduler::process_to_add() {
   LockGuard guard{this->lock_};
   for (auto &it : this->to_add_) {
-    if (it->remove) {
+    if (is_item_removed_(it.get())) {
+      // Recycle cancelled items
+      this->recycle_item_(std::move(it));
       continue;
     }
 
@@ -525,6 +565,10 @@ size_t HOT Scheduler::cleanup_() {
 }
 void HOT Scheduler::pop_raw_() {
   std::pop_heap(this->items_.begin(), this->items_.end(), SchedulerItem::cmp);
+
+  // Instead of destroying, recycle the item
+  this->recycle_item_(std::move(this->items_.back()));
+
   this->items_.pop_back();
 }
 
@@ -559,7 +603,7 @@ bool HOT Scheduler::cancel_item_locked_(Component *component, const char *name_c
 
   // Check all containers for matching items
 #ifndef ESPHOME_THREAD_SINGLE
-  // Only check defer queue for timeouts (intervals never go there)
+  // Mark items in defer queue as cancelled (they'll be skipped when processed)
   if (type == SchedulerItem::TIMEOUT) {
     for (auto &item : this->defer_queue_) {
       if (this->matches_item_(item, component, name_cstr, type, match_retry)) {
@@ -571,11 +615,22 @@ bool HOT Scheduler::cancel_item_locked_(Component *component, const char *name_c
 #endif /* not ESPHOME_THREAD_SINGLE */
 
   // Cancel items in the main heap
-  for (auto &item : this->items_) {
-    if (this->matches_item_(item, component, name_cstr, type, match_retry)) {
-      this->mark_item_removed_(item.get());
+  // Special case: if the last item in the heap matches, we can remove it immediately
+  // (removing the last element doesn't break heap structure)
+  if (!this->items_.empty()) {
+    auto &last_item = this->items_.back();
+    if (this->matches_item_(last_item, component, name_cstr, type, match_retry)) {
+      this->recycle_item_(std::move(this->items_.back()));
+      this->items_.pop_back();
       total_cancelled++;
-      this->to_remove_++;  // Track removals for heap items
+    }
+    // For other items in heap, we can only mark for removal (can't remove from middle of heap)
+    for (auto &item : this->items_) {
+      if (this->matches_item_(item, component, name_cstr, type, match_retry)) {
+        this->mark_item_removed_(item.get());
+        total_cancelled++;
+        this->to_remove_++;  // Track removals for heap items
+      }
     }
   }
 
@@ -754,4 +809,25 @@ bool HOT Scheduler::SchedulerItem::cmp(const std::unique_ptr<SchedulerItem> &a,
   return a->next_execution_ > b->next_execution_;
 }
 
+void Scheduler::recycle_item_(std::unique_ptr<SchedulerItem> item) {
+  if (!item)
+    return;
+
+  if (this->scheduler_item_pool_.size() < MAX_POOL_SIZE) {
+    // Clear callback to release captured resources
+    item->callback = nullptr;
+    // Clear dynamic name if any
+    item->clear_dynamic_name();
+    this->scheduler_item_pool_.push_back(std::move(item));
+#ifdef ESPHOME_DEBUG_SCHEDULER
+    ESP_LOGD(TAG, "Recycled item to pool (pool size now: %zu)", this->scheduler_item_pool_.size());
+#endif
+  } else {
+#ifdef ESPHOME_DEBUG_SCHEDULER
+    ESP_LOGD(TAG, "Pool full (size: %zu), deleting item", this->scheduler_item_pool_.size());
+#endif
+  }
+  // else: unique_ptr will delete the item when it goes out of scope
+}
+
 }  // namespace esphome

esphome/core/scheduler.h

@@ -142,11 +142,7 @@ class Scheduler {
     }
 
     // Destructor to clean up dynamic names
-    ~SchedulerItem() {
-      if (name_is_dynamic) {
-        delete[] name_.dynamic_name;
-      }
-    }
+    ~SchedulerItem() { clear_dynamic_name(); }
 
     // Delete copy operations to prevent accidental copies
     SchedulerItem(const SchedulerItem &) = delete;
@@ -159,13 +155,19 @@ class Scheduler {
     // Helper to get the name regardless of storage type
     const char *get_name() const { return name_is_dynamic ? name_.dynamic_name : name_.static_name; }
 
+    // Helper to clear dynamic name if allocated
+    void clear_dynamic_name() {
+      if (name_is_dynamic && name_.dynamic_name) {
+        delete[] name_.dynamic_name;
+        name_.dynamic_name = nullptr;
+        name_is_dynamic = false;
+      }
+    }
+
     // Helper to set name with proper ownership
     void set_name(const char *name, bool make_copy = false) {
       // Clean up old dynamic name if any
-      if (name_is_dynamic && name_.dynamic_name) {
-        delete[] name_.dynamic_name;
-        name_is_dynamic = false;
-      }
+      clear_dynamic_name();
 
       if (!name) {
         // nullptr case - no name provided
@@ -214,6 +216,15 @@ class Scheduler {
   // Common implementation for cancel operations
   bool cancel_item_(Component *component, bool is_static_string, const void *name_ptr, SchedulerItem::Type type);
 
+  // Helper to check if two scheduler item names match
+  inline bool HOT names_match_(const char *name1, const char *name2) const {
+    // Check pointer equality first (common for static strings), then string contents
+    // The core ESPHome codebase uses static strings (const char*) for component names,
+    // making pointer comparison effective. The std::string overloads exist only for
+    // compatibility with external components but are rarely used in practice.
+    return (name1 != nullptr && name2 != nullptr) && ((name1 == name2) || (strcmp(name1, name2) == 0));
+  }
+
   // Helper function to check if item matches criteria for cancellation
   inline bool HOT matches_item_(const std::unique_ptr<SchedulerItem> &item, Component *component, const char *name_cstr,
                                 SchedulerItem::Type type, bool match_retry, bool skip_removed = true) const {
@@ -221,29 +232,20 @@ class Scheduler {
         (match_retry && !item->is_retry)) {
       return false;
     }
-    const char *item_name = item->get_name();
-    if (item_name == nullptr) {
-      return false;
-    }
-    // Fast path: if pointers are equal
-    // This is effective because the core ESPHome codebase uses static strings (const char*)
-    // for component names. The std::string overloads exist only for compatibility with
-    // external components, but are rarely used in practice.
-    if (item_name == name_cstr) {
-      return true;
-    }
-    // Slow path: compare string contents
-    return strcmp(name_cstr, item_name) == 0;
+    return this->names_match_(item->get_name(), name_cstr);
   }
 
   // Helper to execute a scheduler item
   void execute_item_(SchedulerItem *item, uint32_t now);
 
   // Helper to check if item should be skipped
-  bool should_skip_item_(const SchedulerItem *item) const {
-    return item->remove || (item->component != nullptr && item->component->is_failed());
+  bool should_skip_item_(SchedulerItem *item) const {
+    return is_item_removed_(item) || (item->component != nullptr && item->component->is_failed());
   }
 
+  // Helper to recycle a SchedulerItem
+  void recycle_item_(std::unique_ptr<SchedulerItem> item);
+
   // Helper to check if item is marked for removal (platform-specific)
   // Returns true if item should be skipped, handles platform-specific synchronization
   // For ESPHOME_THREAD_MULTI_NO_ATOMICS platforms, the caller must hold the scheduler lock before calling this
@@ -280,8 +282,9 @@ class Scheduler {
   bool has_cancelled_timeout_in_container_(const Container &container, Component *component, const char *name_cstr,
                                            bool match_retry) const {
     for (const auto &item : container) {
-      if (item->remove && this->matches_item_(item, component, name_cstr, SchedulerItem::TIMEOUT, match_retry,
-                                              /* skip_removed= */ false)) {
+      if (is_item_removed_(item.get()) &&
+          this->matches_item_(item, component, name_cstr, SchedulerItem::TIMEOUT, match_retry,
+                              /* skip_removed= */ false)) {
         return true;
       }
     }
@@ -297,6 +300,16 @@ class Scheduler {
 #endif                                                      /* ESPHOME_THREAD_SINGLE */
   uint32_t to_remove_{0};
 
+  // Memory pool for recycling SchedulerItem objects to reduce heap churn.
+  // Design decisions:
+  // - std::vector is used instead of a fixed array because many systems only need 1-2 scheduler items
+  // - The vector grows dynamically up to MAX_POOL_SIZE (5) only when needed, saving memory on simple setups
+  // - Pool size of 5 matches typical usage (2-4 timers) while keeping memory overhead low (~250 bytes on ESP32)
+  // - The pool significantly reduces heap fragmentation which is critical because heap allocation/deallocation
+  //   can stall the entire system, causing timing issues and dropped events for any components that need
+  //   to synchronize between tasks (see https://github.com/esphome/backlog/issues/52)
+  std::vector<std::unique_ptr<SchedulerItem>> scheduler_item_pool_;
+
 #ifdef ESPHOME_THREAD_MULTI_ATOMICS
   /*
    * Multi-threaded platforms with atomic support: last_millis_ needs atomic for lock-free updates