vfs/SetAttr: async throttle setattr storms

[Xuyuchao-juice]

Background:
When mounting with -o writeback_cache and running sequential writes, the kernel issues excessive setattr (time updates), which drags down throughput.

Change:
Add a short-lived attr cache plus an async SetAttr worker pool so pure time updates can be deduplicated, rate-limited, and handled without blocking the foreground path.

Test Method and Script:
Use 4K block sequential writing to create 4 files each of 512M size.
meta：sqlite

// seq_write.go
package main

import (
	"crypto/rand"
	"fmt"
	"os"
	"time"
)

const (
	blockSize = 4 * 1024          // 4KB
	fileSize  = 512 * 1024 * 1024 // 512MB
)

func main() {
	numIterations := 4

	// 记录总耗时
	totalDuration := time.Duration(0)

	for iteration := 1; iteration <= numIterations; iteration++ {
		// 计算文件中的块数
		numBlocks := fileSize / blockSize

		// 创建文件并预分配空间
		fileName := fmt.Sprintf("seq_write_%d.dat", iteration)
		file, err := os.Create(fileName)
		if err != nil {
			fmt.Printf("创建文件 %s 失败: %v\n", fileName, err)
			continue
		}

		fmt.Printf("\n开始写入文件 %s (迭代 %d/%d)\n", fileName, iteration, numIterations)

		// 预分配文件空间
		if err := file.Truncate(fileSize); err != nil {
			fmt.Printf("预分配文件空间失败: %v\n", err)
			file.Close()
			continue
		}

		// 生成随机数据块
		dataBlock := make([]byte, blockSize)
		if _, err := rand.Read(dataBlock); err != nil {
			fmt.Printf("生成随机数据失败: %v\n", err)
			file.Close()
			continue
		}

		// 记录开始时间
		startTime := time.Now()

		for writtenBlocks := 1; writtenBlocks <= numBlocks; writtenBlocks++ {
			offset := int64(writtenBlocks-1) * blockSize

			if _, err := file.WriteAt(dataBlock, offset); err != nil {
				fmt.Printf("写入文件失败: %v\n", err)
				break
			}

			if writtenBlocks%100 == 0 || writtenBlocks == numBlocks {
				progress := float64(writtenBlocks) / float64(numBlocks) * 100
				fmt.Printf("\r进度: %.2f%% (%d/%d)", progress, writtenBlocks, numBlocks)
			}
		}
		fmt.Println()

		// 关闭文件
		file.Close()

		// 计算并打印执行时间
		duration := time.Since(startTime)
		totalDuration += duration
		fmt.Printf("文件 %s 写入完成！耗时: %v\n", fileName, duration)
		fmt.Printf("写入速度: %.2f MB/s\n", float64(fileSize)/1024/1024/duration.Seconds())
	}

	// 打印总统计信息
	fmt.Printf("\n所有 %d 个文件写入完成！\n", numIterations)
	fmt.Printf("总耗时: %v\n", totalDuration)
	fmt.Printf("平均每个文件耗时: %v\n", totalDuration/time.Duration(numIterations))
	fmt.Printf("平均写入速度: %.2f MB/s\n", float64(fileSize*numIterations)/1024/1024/totalDuration.Seconds())
}

Test Result:

[Xuyuchao-juice]

看看这几个数值设置是否合理：

1. 目前缓存的有效时间10s
2. 后台任务每1分钟清理一次过期缓存
3. 并发setattr goroutine最多100个

[jiefenghuang]

开了writeback_cache，可以把max-fuse-io调大，应该能减少请求比例。
Increasing max-fuse-io to 1MiB should help reduce the proportion of requests.