l4 evict and swiss api improve

src/Nitload.Common/HighPerformance/Memory/L4.Predict.cs

@@ -16,6 +16,10 @@ namespace Nitload.Common.HighPerformance.Memory
     public unsafe partial class L4
     {
 
+        private int _active_line_count;
+
+        private int _last_evict_pos;
+
         internal partial int Predict()
         {
             if (!TryGetCurrentPredictor(out IL4Predictor? p)) {
@@ -23,12 +27,13 @@ namespace Nitload.Common.HighPerformance.Memory
             }
             int total_count = 0, count;
             Span<L4.AddressSource> as_sp = stackalloc L4.AddressSource[4];
+            SwissEnumerator e = _addressMap.GetWriteBufferEnumerator();
             Predict:
             while ((count = p.PredictMission(in as_sp)) > 0)
             {
                 for (int i = 0; i < count; i++)
                 {
-                    if (_addressMap.TryGet(as_sp[i], out ulong map_handle_adr))
+                    if (_addressMap.TryGetWrite(as_sp[i], out ulong map_handle_adr))
                     {
                         L4.CacheLineMetadata* map_handle = (L4.CacheLineMetadata*)map_handle_adr;
                         map_handle->IncreasePredictCount();
@@ -38,13 +43,31 @@ namespace Nitload.Common.HighPerformance.Memory
                         uint size = as_sp[i].Size;
                         int handle_idx = _bufferAllocator.Alloc(size);
                         L4.CacheLineMetadata* handle = _bufferAllocator.MetadataAt(handle_idx);
+                        p.Prefetch(as_sp[i], handle->Position, handle->SizeMask);
                         if (!_addressMap.TryAddOrSet(as_sp[i], (ulong)handle)) {
                             throw new InvalidOperationException();
                         }
+                        _active_line_count++;
+
+                        if (_active_line_count >= MaxActiveCacheLineCount)
+                        {
+                            int retry = 0;
+                            const int max_retry = 3;
+                            for (; retry < max_retry && EvictOnce(ref e); retry++) {
+                                e = _addressMap.GetWriteBufferEnumerator(_last_evict_pos);
+                            }
+                            if (retry >= max_retry) {
+                                throw new OverflowException("Evict failed, maybe all cache lines are in use.");
+                            }
+                            _active_line_count--;
+                        }
                     }
                 }
                 total_count += count;
-                if (total_count >= 2048) {
+                if (total_count >= 2048)
+                {
+                    _ = Interlocked.Add(ref _predict_count, total_count);
+                    _last_evict_pos = e.CurrentOffset;
                     return total_count;
                 }
             }
@@ -53,22 +76,36 @@ namespace Nitload.Common.HighPerformance.Memory
                 _ = TryGetCurrentPredictor(out p);
                 goto Predict;
             }
+            _ = Interlocked.Add(ref _predict_count, total_count);
+            _last_evict_pos = e.CurrentOffset;
             return total_count;
         }
 
+        private bool EvictOnce(
+                     ref SwissEnumerator e
+        )
+        {
+            while (e.MoveNext())
+            {
+                (AddressSource source, ulong cache_head_ptr) = e.Current;
+                L4.CacheLineMetadata* handle = (L4.CacheLineMetadata*)cache_head_ptr;
+                if (Volatile.Read(ref handle->ConsumeCount) == Volatile.Read(ref handle->ConsumeCount))
+                {
+                    // evict
+                    _ = _addressMap.TryRemove(source, out _);
+                    _bufferAllocator.Free(handle);
+                    return true;
+                }
+            }
+            return false;
+        }
+
     }
 }
 
-/*
-goto 本身没问题，但当前控制流有几个需要先确认的点：
-
-高：换 predictor 的分支看起来反了。 MoveNextPredictor() 在 _predict 到达 _enter 时返回 true。因此 L4.Predict.cs:51 现在会在队列耗尽时再次取 predictor；TryGetCurrentPredictor 失败后仍 goto Predict，接着会用空的 p 调 PredictMission。反过来，有下一个 predictor 时它会提前返回，不再处理后续 predictor。
-
-高：新分配的 metadata 没有增加预测计数。 现有映射分支会调用 IncreasePredictCount()，但新分配分支只分配并写入 Swiss。若该计数表示未消费预测数，新条目也需要相同的计数处理；还要确认 TLSF 重用 metadata 时会重置旧计数。
-
-高：同一轮内重复的 AddressSource 可能重复分配。 SwissTable.TryGet 只查 published buffer，而新条目先写入 inactive buffer。后续再次预测相同 key 时会查不到，重新分配 metadata，再由 TryAddOrSet 覆盖先前 value，旧分配就失去映射。
-
-中：需要校验 PredictMission 的返回值范围。 Span 容量是 4；count > 4 会让 as_sp[i] 越界，负值则被当作本轮结束。建议契约明确要求 0..4，并在 L4 边界验证。
-
-本轮未改代码。L4.Predict.cs 的文件诊断通过，但尚未运行完整构建或运行时测试。
-*/
+// Eviction follow-up:
+// 1. Compare ConsumeCount with PredictCount; reset both on TLSF slot reuse and count new predictions.
+// 2. Fix retry/result handling; decrement _active_line_count only after a successful eviction.
+// 3. Correct TLSF reverse-slab pointer validation before freeing metadata pointers.
+// 4. Defer TLSF Free until retired published buffers and outstanding handles can no longer reference the line.
+// 5. Add Swiss TryRemoveAt(offset) to tombstone the enumerated slot without hashing and probing again.