feat(mm): page reclaim for file-backed memory pressure #804
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -6,7 +6,7 @@ use alloc::{ | |
| use core::{ | ||
| num::NonZeroUsize, | ||
| ops::Range, | ||
| sync::atomic::{AtomicBool, AtomicU8, Ordering}, | ||
| task::Context, | ||
| }; | ||
|
|
||
|
|
@@ -375,23 +375,90 @@ intrusive_adapter!(EvictListenerAdapter = Box<EvictListener>: EvictListener { li | |
|
|
||
| struct CachedFileShared { | ||
| page_cache: Mutex<LruCache<u32, PageCache>>, | ||
| evict_listeners: spin::Mutex<LinkedList<EvictListenerAdapter>>, | ||
| } | ||
|
|
||
| impl CachedFileShared { | ||
| pub fn new() -> Self { | ||
| Self { | ||
| page_cache: Mutex::new(LruCache::new(NonZeroUsize::new(64).unwrap())), | ||
| evict_listeners: spin::Mutex::new(LinkedList::default()), | ||
| } | ||
| } | ||
|
|
||
| #[allow(dead_code)] | ||
| pub fn new_unbounded() -> Self { | ||
| Self { | ||
| page_cache: Mutex::new(LruCache::unbounded()), | ||
| evict_listeners: spin::Mutex::new(LinkedList::default()), | ||
| } | ||
| } | ||
|
|
||
| fn try_evict_clean_pages(&self, max: usize) -> usize { | ||
| let Some(mut cache) = self.page_cache.try_lock() else { | ||
| return 0; | ||
| }; | ||
| let mut to_evict = alloc::vec::Vec::new(); | ||
| for (&pn, page) in cache.iter() { | ||
| if !page.dirty && to_evict.len() < max { | ||
| to_evict.push(pn); | ||
| } | ||
| } | ||
|
Contributor
There was a problem hiding this comment. 此处 |
||
| let count = to_evict.len(); | ||
|
Contributor
There was a problem hiding this comment.
当前：cache 锁 → 建议分离 pop 和 notify 阶段：先在 cache 锁内收集要驱逐的页面并 pop 出来（已脱离 cache），释放 cache 锁后再通知 listener： fn try_evict_clean_pages(&self, max: usize) -> usize {
let Some(mut cache) = self.page_cache.try_lock() else { return 0; };
let mut evicted_pages: alloc::vec::Vec<(u32, PageCache)> = alloc::vec::Vec::new();
let clean_pns: alloc::vec::Vec<u32> = cache.iter()
.filter(|(_, p)| !p.dirty).take(max).map(|(&pn, _)| pn).collect();
for pn in clean_pns {
if let Some(page) = cache.pop(&pn) {
evicted_pages.push((pn, page));
}
}
let count = evicted_pages.len();
drop(cache); // 释放 cache 锁后再通知 listener
for (pn, page) in &evicted_pages {
for listener in self.evict_listeners.lock().iter() {
(listener.listener)(*pn, page);
}
}
count
}这样 listener 回调可以安全地访问 page_cache 而不会死锁。 |
||
| for pn in to_evict { | ||
| if let Some(page) = cache.pop(&pn) { | ||
| for listener in self.evict_listeners.lock().iter() { | ||
|
Contributor
There was a problem hiding this comment. 阻塞性问题：回收路径中调用阻塞锁
如果另一个 CPU 正持有 commit 4a9e883 的提交信息说「Evict listeners registered by the address-space backend can call blocking functions that sleep」，但这个理由只说明了监听器回调内部的行为，并未解决回收路径在原子上下文中调用阻塞锁的问题。 建议修复方案：
Contributor
Author
There was a problem hiding this comment. This concern is not actually a blocking issue. Analysis:
|
||
| (listener.listener)(pn, &page); | ||
| } | ||
| } | ||
| } | ||
|
Contributor
There was a problem hiding this comment. 锁序文档写得很好，明确了 |
||
| count | ||
| } | ||
|
Contributor
There was a problem hiding this comment. 使用 |
||
| } | ||
|
|
||
| static GLOBAL_CACHED_FILES: spin::RwLock<alloc::vec::Vec<Weak<CachedFileShared>>> = | ||
| spin::RwLock::new(alloc::vec::Vec::new()); | ||
|
Comment on lines
+450
to
+451
Contributor
There was a problem hiding this comment. 改进建议（非阻塞）： 建议后续迭代中考虑：
当前实现在大多数场景下可以正常工作（RECLAIM_IN_PROGRESS 防止递归），作为首版实现可以接受，但建议作为后续优化点记录。 |
||
|
|
||
| static RECLAIM_IN_PROGRESS: AtomicBool = AtomicBool::new(false); | ||
|
|
||
| pub fn page_cache_reclaim(num_pages: usize) -> usize { | ||
|
|
||
| if RECLAIM_IN_PROGRESS.swap(true, Ordering::Acquire) { | ||
|
|
||
| return 0; | ||
| } | ||
|
|
||
| let mut reclaimed = 0; | ||
| let Some(mut guard) = GLOBAL_CACHED_FILES.try_write() else { | ||
| RECLAIM_IN_PROGRESS.store(false, Ordering::Release); | ||
| return 0; | ||
| }; | ||
| let files: alloc::vec::Vec<Arc<CachedFileShared>> = | ||
|
Contributor
There was a problem hiding this comment. 这里两次调用 let (alive, dead): (Vec<_>, Vec<_>) = guard.drain(..).partition(|w| w.upgrade().is_some());
*guard = alive;或者用 |
||
| guard.iter().filter_map(|w| w.upgrade()).collect(); | ||
|
Contributor
There was a problem hiding this comment. 阻塞：回收路径中的堆分配在内存压力下可能失败。 page_cache_reclaim 的设计目的是在内存不足时被调用，但其实现包含多处堆分配：
在内存压力下，这些 Vec 分配本身可能失败或再次触发回收，导致无限递归或 OOM。 建议修复：避免堆分配——例如用 try_write 锁遍历 Weak 列表逐个 upgrade 后立即驱逐，使用有界栈上缓冲区或直接在 LRU 迭代时驱逐。
Contributor
There was a problem hiding this comment. 非阻塞建议：此处 可考虑使用固定大小的栈缓冲区（类似 |
||
| guard.retain(|w| w.upgrade().is_some()); | ||
| drop(guard); | ||
|
|
||
| let target = num_pages.max(64) * 4; | ||
| for file in &files { | ||
| let freed = file.try_evict_clean_pages(target - reclaimed); | ||
|
Contributor
There was a problem hiding this comment. 回收路径中的堆分配与内存压力场景矛盾。
如果堆内存已耗尽，这些分配可能失败或递归进入 reclaim。建议：
例如： let guard = GLOBAL_CACHED_FILES.read();
for weak in guard.iter() {
if let Some(file) = weak.upgrade() {
reclaimed += file.try_evict_clean_pages(target - reclaimed);
if reclaimed >= target { break; }
}
} |
||
| reclaimed += freed; | ||
| if reclaimed >= target { | ||
| break; | ||
| } | ||
| } | ||
|
|
||
| if reclaimed > 0 { | ||
| debug!( | ||
| "page_cache_reclaim: evicted {} clean pages across {} files", | ||
| reclaimed, | ||
| files.len() | ||
| ); | ||
| } | ||
|
|
||
| RECLAIM_IN_PROGRESS.store(false, Ordering::Release); | ||
|
|
||
| reclaimed | ||
| } | ||
|
|
||
| fn register_cached_file(file: &Arc<CachedFileShared>) { | ||
| GLOBAL_CACHED_FILES.write().push(Arc::downgrade(file)); | ||
| } | ||
|
Comment on lines
+490
to
494
|
||
|
|
||
| /// A file handle with an LRU page cache for buffered I/O. | ||
|
|
@@ -435,22 +502,27 @@ impl CachedFile { | |
| let in_memory = location.filesystem().name() == "tmpfs"; | ||
|
|
||
| let mut guard = location.user_data(); | ||
| let (shared, is_new) = | ||
| if let Some(shared) = guard.get::<FileUserData>().and_then(|it| it.get()) { | ||
| (shared, false) | ||
| } else { | ||
| let (shared, user_data) = if in_memory { | ||
| let shared = Arc::new(CachedFileShared::new_unbounded()); | ||
| (shared.clone(), FileUserData::Strong(shared)) | ||
| } else { | ||
| let shared = Arc::new(CachedFileShared::new()); | ||
| let user_data = FileUserData::Weak(Arc::downgrade(&shared)); | ||
| (shared, user_data) | ||
| }; | ||
| guard.insert(user_data); | ||
| (shared, true) | ||
| }; | ||
| drop(guard); | ||
|
|
||
| if is_new { | ||
| register_cached_file(&shared); | ||
| } | ||
|
|
||
| Self { | ||
| inner: location, | ||
| shared, | ||
|
|
||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
try_page_reclaim()未被分配器调用（见 axalloc/lib.rs 的 inline comment），此处注册的回调实际上是死代码。需先完成分配器后端集成。