sascha/esphome
1
0
Fork 0
mirror of https://github.com/esphome/esphome.git synced 2025-10-31 15:12:06 +00:00
Code Issues Releases Wiki Activity

core/scheduler: Make millis_64_ rollover monotonic on SMP

Browse Source 
The current implementation uses only memory_order_relaxed on all atomic
accesses. That protects each variable individually but not the semantic
link between the low word (last_millis_) and the high-word epoch counter
(millis_major_). On a multi-core target a reader could observe a freshly
stored low word before seeing the matching increment of the epoch,
causing a ~49-day negative jump.

Key fixes
- Release/acquire pairing
  - writer: compare_exchange_weak(..., memory_order_release, …)
  - reader: first load of last_millis_ now uses memory_order_acquire
  - ensures any core that sees the new low word also sees the updated
    high word
- Epoch-coherency retry loop
  - re-loads millis_major_ after the update and retries if it changed,
    guaranteeing monotonicity even when another core rolls over
    concurrently
- millis_major_ promoted to std::atomic<uint16_t> on SMP platforms
  - removes the formal data race at negligible cost
- new macros for better readability
  - ESPHOME_SINGLE_CORE – currently ESP8266/RP2040 only
  - ESPHOME_ATOMIC_SCHEDULER – all others except LibreTiny
- Logging and comments
  - loads atomics safely in debug output
  - updated inline docs to match the memory ordering

Behavior on single-core or non-atomic platforms is unchanged; multi-core
targets now get a provably monotonic 64-bit millisecond clock with
minimal overhead.
This commit is contained in:
RubenKelevra
2025-07-18 08:12:00 +02:00
parent 89b9bddf1b
commit 211739bba0
3 changed files with 136 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
esphome/core/defines.h
View File

@@ -229,6 +229,16 @@
#define USE_SOCKET_SELECT_SUPPORT
#endif
// Helper macro for platforms that lack atomic scheduler support
#if defined(USE_ESP8266) || defined(USE_RP2040)
#define ESPHOME_SINGLE_CORE
#endif
// Helper macro for platforms with atomic scheduler support
#if !defined(ESPHOME_SINGLE_CORE) && !defined(USE_LIBRETINY)
#define ESPHOME_ATOMIC_SCHEDULER
#endif
// Disabled feature flags
// #define USE_BSEC // Requires a library with proprietary license
// #define USE_BSEC2 // Requires a library with proprietary license

166
esphome/core/scheduler.cpp
View File

@@ -54,7 +54,7 @@ static void validate_static_string(const char *name) {
ESP_LOGW(TAG, "WARNING: Scheduler name '%s' at %p might be on heap (static ref at %p)", name, name, static_str);
}
}
#endif
#endif /* ESPHOME_DEBUG_SCHEDULER */
// A note on locking: the `lock_` lock protects the `items_` and `to_add_` containers. It must be taken when writing to
// them (i.e. when adding/removing items, but not when changing items). As items are only deleted from the loop task,
@@ -82,9 +82,9 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
item->callback = std::move(func);
item->remove = false;
#if !defined(USE_ESP8266) && !defined(USE_RP2040)
#ifndef ESPHOME_SINGLE_CORE
// Special handling for defer() (delay = 0, type = TIMEOUT)
// ESP8266 and RP2040 are excluded because they don't need thread-safe defer handling
// 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_};
@@ -92,7 +92,7 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
this->defer_queue_.push_back(std::move(item));
return;
}
#endif
#endif /* not ESPHOME_SINGLE_CORE */
// Get fresh timestamp for new timer/interval - ensures accurate scheduling
const auto now = this->millis_64_(millis()); // Fresh millis() call
@@ -123,7 +123,7 @@ void HOT Scheduler::set_timer_common_(Component *component, SchedulerItem::Type
ESP_LOGD(TAG, "set_%s(name='%s/%s', %s=%" PRIu32 ", offset=%" PRIu32 ")", type_str, item->get_source(),
name_cstr ? name_cstr : "(null)", type_str, delay, static_cast<uint32_t>(item->next_execution_ - now));
}
#endif
#endif /* ESPHOME_DEBUG_SCHEDULER */
LockGuard guard{this->lock_};
// If name is provided, do atomic cancel-and-add
@@ -231,7 +231,7 @@ optional<uint32_t> HOT Scheduler::next_schedule_in(uint32_t now) {
return item->next_execution_ - now_64;
}
void HOT Scheduler::call(uint32_t now) {
#if !defined(USE_ESP8266) && !defined(USE_RP2040)
#ifndef ESPHOME_SINGLE_CORE
// Process defer queue first to guarantee FIFO execution order for deferred items.
// Previously, defer() used the heap which gave undefined order for equal timestamps,
// causing race conditions on multi-core systems (ESP32, BK7200).
@@ -239,8 +239,7 @@ void HOT Scheduler::call(uint32_t now) {
// - Deferred items (delay=0) go directly to defer_queue_ in set_timer_common_
// - Items execute in exact order they were deferred (FIFO guarantee)
// - No deferred items exist in to_add_, so processing order doesn't affect correctness
// ESP8266 and RP2040 don't use this queue - they fall back to the heap-based approach
// (ESP8266: single-core, RP2040: empty mutex implementation).
// Single-core platforms don't use this queue and fall back to the heap-based approach.
//
// Note: Items cancelled via cancel_item_locked_() are marked with remove=true but still
// processed here. They are removed from the queue normally via pop_front() but skipped
@@ -262,7 +261,7 @@ void HOT Scheduler::call(uint32_t now) {
this->execute_item_(item.get(), now);
}
}
#endif
#endif /* not ESPHOME_SINGLE_CORE */
// Convert the fresh timestamp from main loop to 64-bit for scheduler operations
const auto now_64 = this->millis_64_(now); // 'now' from parameter - fresh from Application::loop()
@@ -274,13 +273,15 @@ void HOT Scheduler::call(uint32_t now) {
if (now_64 - last_print > 2000) {
last_print = now_64;
std::vector<std::unique_ptr<SchedulerItem>> old_items;
#if !defined(USE_ESP8266) && !defined(USE_RP2040) && !defined(USE_LIBRETINY)
ESP_LOGD(TAG, "Items: count=%zu, now=%" PRIu64 " (%u, %" PRIu32 ")", this->items_.size(), now_64,
this->millis_major_, this->last_millis_.load(std::memory_order_relaxed));
#else
ESP_LOGD(TAG, "Items: count=%zu, now=%" PRIu64 " (%u, %" PRIu32 ")", this->items_.size(), now_64,
#ifdef ESPHOME_ATOMIC_SCHEDULER
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);
#else /* not ESPHOME_ATOMIC_SCHEDULER */
ESP_LOGD(TAG, "Items: count=%zu, now=%" PRIu64 " (%" PRIu16 ", %" PRIu32 ")", this->items_.size(), now_64,
this->millis_major_, this->last_millis_);
#endif
#endif /* else ESPHOME_ATOMIC_SCHEDULER */
while (!this->empty_()) {
std::unique_ptr<SchedulerItem> item;
{
@@ -305,7 +306,7 @@ void HOT Scheduler::call(uint32_t now) {
std::make_heap(this->items_.begin(), this->items_.end(), SchedulerItem::cmp);
}
}
#endif // ESPHOME_DEBUG_SCHEDULER
#endif /* ESPHOME_DEBUG_SCHEDULER */
// If we have too many items to remove
if (this->to_remove_ > MAX_LOGICALLY_DELETED_ITEMS) {
@@ -352,7 +353,7 @@ void HOT Scheduler::call(uint32_t now) {
ESP_LOGV(TAG, "Running %s '%s/%s' with interval=%" PRIu32 " next_execution=%" PRIu64 " (now=%" PRIu64 ")",
item->get_type_str(), item->get_source(), item_name ? item_name : "(null)", item->interval,
item->next_execution_, now_64);
#endif
#endif /* ESPHOME_DEBUG_SCHEDULER */
// Warning: During callback(), a lot of stuff can happen, including:
// - timeouts/intervals get added, potentially invalidating vector pointers
@@ -460,7 +461,7 @@ bool HOT Scheduler::cancel_item_locked_(Component *component, const char *name_c
size_t total_cancelled = 0;
// Check all containers for matching items
#if !defined(USE_ESP8266) && !defined(USE_RP2040)
#ifndef ESPHOME_SINGLE_CORE
// Only check defer queue for timeouts (intervals never go there)
if (type == SchedulerItem::TIMEOUT) {
for (auto &item : this->defer_queue_) {
@@ -470,7 +471,7 @@ bool HOT Scheduler::cancel_item_locked_(Component *component, const char *name_c
}
}
}
#endif
#endif /* not ESPHOME_SINGLE_CORE */
// Cancel items in the main heap
for (auto &item : this->items_) {
@@ -496,7 +497,7 @@ bool HOT Scheduler::cancel_item_locked_(Component *component, const char *name_c
uint64_t Scheduler::millis_64_(uint32_t now) {
// THREAD SAFETY NOTE:
// This function can be called from multiple threads simultaneously on ESP32/LibreTiny.
// On single-threaded platforms (ESP8266, RP2040), atomics are not needed.
// On single-core platforms, atomics are not needed.
//
// IMPORTANT: Always pass fresh millis() values to this function. The implementation
// handles out-of-order timestamps between threads, but minimizing time differences
@@ -508,99 +509,128 @@ uint64_t Scheduler::millis_64_(uint32_t now) {
// This prevents race conditions at the rollover boundary without requiring
// 64-bit atomics or locking on every call.
#ifdef ESPHOME_ATOMIC_SCHEDULER
for (;;) {
uint16_t major = this->millis_major_.load(std::memory_order_acquire);
#else /* not ESPHOME_ATOMIC_SCHEDULER */
uint16_t major = this->millis_major_;
#endif /* else ESPHOME_ATOMIC_SCHEDULER */
#ifdef USE_LIBRETINY
// LibreTiny: Multi-threaded but lacks atomic operation support
// TODO: If LibreTiny ever adds atomic support, remove this entire block and
// let it fall through to the atomic-based implementation below
// We need to use a lock when near the rollover boundary to prevent races
uint32_t last = this->last_millis_;
// LibreTiny: Multi-threaded but lacks atomic operation support
// TODO: If LibreTiny ever adds atomic support, remove this entire block and
// let it fall through to the atomic-based implementation below
// We need to use a lock when near the rollover boundary to prevent races
uint32_t last = this->last_millis_;
// Define a safe window around the rollover point (10 seconds)
// This covers any reasonable scheduler delays or thread preemption
static const uint32_t ROLLOVER_WINDOW = 10000; // 10 seconds in milliseconds
// Define a safe window around the rollover point (10 seconds)
// This covers any reasonable scheduler delays or thread preemption
static const uint32_t ROLLOVER_WINDOW = 10000; // 10 seconds in milliseconds
// Check if we're near the rollover boundary (close to std::numeric_limits<uint32_t>::max() or just past 0)
bool near_rollover = (last > (std::numeric_limits<uint32_t>::max() - ROLLOVER_WINDOW)) || (now < ROLLOVER_WINDOW);
// Check if we're near the rollover boundary (close to std::numeric_limits<uint32_t>::max() or just past 0)
bool near_rollover = (last > (std::numeric_limits<uint32_t>::max() - ROLLOVER_WINDOW)) || (now < ROLLOVER_WINDOW);
if (near_rollover || (now < last && (last - now) > HALF_MAX_UINT32)) {
// Near rollover or detected a rollover - need lock for safety
LockGuard guard{this->lock_};
// Re-read with lock held
last = this->last_millis_;
if (near_rollover || (now < last && (last - now) > HALF_MAX_UINT32)) {
// Near rollover or detected a rollover - need lock for safety
LockGuard guard{this->lock_};
// Re-read with lock held
last = this->last_millis_;
if (now < last && (last - now) > HALF_MAX_UINT32) {
// True rollover detected (happens every ~49.7 days)
this->millis_major_++;
if (now < last && (last - now) > HALF_MAX_UINT32) {
// True rollover detected (happens every ~49.7 days)
this->millis_major_++;
major++;
#ifdef ESPHOME_DEBUG_SCHEDULER
ESP_LOGD(TAG, "Detected true 32-bit rollover at %" PRIu32 "ms (was %" PRIu32 ")", now, last);
#endif
ESP_LOGD(TAG, "Detected true 32-bit rollover at %" PRIu32 "ms (was %" PRIu32 ")", now, last);
#endif /* ESPHOME_DEBUG_SCHEDULER */
}
// Update last_millis_ while holding lock
this->last_millis_ = now;
} else if (now > last) {
// Normal case: Not near rollover and time moved forward
// Update without lock. While this may cause minor races (microseconds of
// backwards time movement), they're acceptable because:
// 1. The scheduler operates at millisecond resolution, not microsecond
// 2. We've already prevented the critical rollover race condition
// 3. Any backwards movement is orders of magnitude smaller than scheduler delays
this->last_millis_ = now;
}
// Update last_millis_ while holding lock
this->last_millis_ = now;
} else if (now > last) {
// Normal case: Not near rollover and time moved forward
// Update without lock. While this may cause minor races (microseconds of
// backwards time movement), they're acceptable because:
// 1. The scheduler operates at millisecond resolution, not microsecond
// 2. We've already prevented the critical rollover race condition
// 3. Any backwards movement is orders of magnitude smaller than scheduler delays
this->last_millis_ = now;
}
// If now <= last and we're not near rollover, don't update
// This minimizes backwards time movement
// If now <= last and we're not near rollover, don't update
// This minimizes backwards time movement
#elif !defined(USE_ESP8266) && !defined(USE_RP2040)
// Multi-threaded platforms with atomic support (ESP32)
uint32_t last = this->last_millis_.load(std::memory_order_relaxed);
#elif defined(ESPHOME_ATOMIC_SCHEDULER)
/*
* Multi-threaded platforms with atomic support (ESP32)
* Acquire so that if we later decide **not** to take the lock we still
* observe a `millis_major_` value coherent with the loaded `last_millis_`.
* The acquire load ensures any later read of `millis_major_` sees its
* corresponding increment.
*/
uint32_t last = this->last_millis_.load(std::memory_order_acquire);
// If we might be near a rollover (large backwards jump), take the lock for the entire operation
// This ensures rollover detection and last_millis_ update are atomic together
if (now < last && (last - now) > HALF_MAX_UINT32) {
// Potential rollover - need lock for atomic rollover detection + update
LockGuard guard{this->lock_};
// Re-read with lock held
// Re-read with lock held; mutex already provides ordering
last = this->last_millis_.load(std::memory_order_relaxed);
if (now < last && (last - now) > HALF_MAX_UINT32) {
// True rollover detected (happens every ~49.7 days)
this->millis_major_++;
this->millis_major_.fetch_add(1, std::memory_order_relaxed);
major++;
#ifdef ESPHOME_DEBUG_SCHEDULER
ESP_LOGD(TAG, "Detected true 32-bit rollover at %" PRIu32 "ms (was %" PRIu32 ")", now, last);
#endif
#endif /* ESPHOME_DEBUG_SCHEDULER */
}
// Update last_millis_ while holding lock to prevent races
this->last_millis_.store(now, std::memory_order_relaxed);
/*
* Update last_millis_ while holding the lock to prevent races
* Publish the new low-word *after* bumping `millis_major_` (done above)
* so readers never see a mismatched pair.
*/
this->last_millis_.store(now, std::memory_order_release);
} else {
// Normal case: Try lock-free update, but only allow forward movement within same epoch
// This prevents accidentally moving backwards across a rollover boundary
while (now > last && (now - last) < HALF_MAX_UINT32) {
if (this->last_millis_.compare_exchange_weak(last, now, std::memory_order_relaxed)) {
if (this->last_millis_.compare_exchange_weak(last, now,
std::memory_order_release, // success
std::memory_order_relaxed)) { // failure
break;
}
// CAS failure means no data was published; relaxed is fine
// last is automatically updated by compare_exchange_weak if it fails
}
}
#else
// Single-threaded platforms (ESP8266, RP2040): No atomics needed
#else /* not USE_LIBRETINY; not ESPHOME_ATOMIC_SCHEDULER */
// Single-core platforms: No atomics needed
uint32_t last = this->last_millis_;
// Check for rollover
if (now < last && (last - now) > HALF_MAX_UINT32) {
this->millis_major_++;
major++;
#ifdef ESPHOME_DEBUG_SCHEDULER
ESP_LOGD(TAG, "Detected true 32-bit rollover at %" PRIu32 "ms (was %" PRIu32 ")", now, last);
#endif
#endif /* ESPHOME_DEBUG_SCHEDULER */
}
// Only update if time moved forward
if (now > last) {
this->last_millis_ = now;
}
#endif
#endif /* else (USE_LIBRETINY / ESPHOME_ATOMIC_SCHEDULER) */
#ifdef ESPHOME_ATOMIC_SCHEDULER
uint16_t major_end = this->millis_major_.load(std::memory_order_relaxed);
if (major_end == major)
return now + (static_cast<uint64_t>(major) << 32);
}
#else /* not ESPHOME_ATOMIC_SCHEDULER */
// Combine major (high 32 bits) and now (low 32 bits) into 64-bit time
return now + (static_cast<uint64_t>(this->millis_major_) << 32);
return now + (static_cast<uint64_t>(major) << 32);
#endif /* ESPHOME_ATOMIC_SCHEDULER */
}
bool HOT Scheduler::SchedulerItem::cmp(const std::unique_ptr<SchedulerItem> &a,

40
esphome/core/scheduler.h
View File

@@ -1,10 +1,11 @@
#pragma once
#include "esphome/core/defines.h"
#include <vector>
#include <memory>
#include <cstring>
#include <deque>
#if !defined(USE_ESP8266) && !defined(USE_RP2040) && !defined(USE_LIBRETINY)
#ifdef ESPHOME_ATOMIC_SCHEDULER
#include <atomic>
#endif
@@ -204,22 +205,37 @@ class Scheduler {
Mutex lock_;
std::vector<std::unique_ptr<SchedulerItem>> items_;
std::vector<std::unique_ptr<SchedulerItem>> to_add_;
#if !defined(USE_ESP8266) && !defined(USE_RP2040)
// ESP8266 and RP2040 don't need the defer queue because:
// ESP8266: Single-core with no preemptive multitasking
// RP2040: Currently has empty mutex implementation in ESPHome
// Both platforms save 40 bytes of RAM by excluding this
#ifndef ESPHOME_SINGLE_CORE
// Single-core platforms don't need the defer queue and save 40 bytes of RAM
std::deque<std::unique_ptr<SchedulerItem>> defer_queue_; // FIFO queue for defer() calls
#endif
#if !defined(USE_ESP8266) && !defined(USE_RP2040) && !defined(USE_LIBRETINY)
// Multi-threaded platforms with atomic support: last_millis_ needs atomic for lock-free updates
#endif /* ESPHOME_SINGLE_CORE */
#ifdef ESPHOME_ATOMIC_SCHEDULER
/*
* Multi-threaded platforms with atomic support: last_millis_ needs atomic for lock-free updates
*
* MEMORY-ORDERING NOTE
* --------------------
* `last_millis_` and `millis_major_` form a single 64-bit timestamp split in half.
* Writers publish `last_millis_` with memory_order_release and readers use
* memory_order_acquire. This ensures that once a reader sees the new low word,
* it also observes the corresponding increment of `millis_major_`.
*/
std::atomic<uint32_t> last_millis_{0};
#else
#else /* not ESPHOME_ATOMIC_SCHEDULER */
// Platforms without atomic support or single-threaded platforms
uint32_t last_millis_{0};
#endif
// millis_major_ is protected by lock when incrementing
#endif /* else ESPHOME_ATOMIC_SCHEDULER */
/*
* Upper 16 bits of the 64-bit millis counter. Incremented only while holding
* `lock_`; read concurrently. Atomic (relaxed) avoids a formal data race.
* Ordering relative to `last_millis_` is provided by its release store and the
* corresponding acquire loads.
*/
#ifdef ESPHOME_ATOMIC_SCHEDULER
std::atomic<uint16_t> millis_major_{0};
#else /* not ESPHOME_ATOMIC_SCHEDULER */
uint16_t millis_major_{0};
#endif /* else ESPHOME_ATOMIC_SCHEDULER */
uint32_t to_remove_{0};
};