cgroups icon indicating copy to clipboard operation
cgroups copied to clipboard

support filter subsystem on `cgroup.Stat()` function

Open yylt opened this issue 6 months ago • 0 comments

When we check metrics using the CRI interface, it queries all subsystems. However, in practice, we might only need CPU and memory information, making the other queries unnecessary and redundant. The metrics interface is often called frequently and periodically.

The call chain is usually ContainerStats -> cgroup.Stat(cntid) or PodStats -> cgroup.Stat(cndids...). For PodStats, this triggers multiple ContainerStats calls.

We've encountered an issue where /proc/partitions is read too frequently which will cause spin locks during the read process.

bt script code:

bpftrace -e '
BEGIN {
    @start_time = nsecs;
    printf("Tracing container processes opening /proc/partitions... (Start Time: %d)\n", @start_time);
}

kprobe:vfs_open
/ strncmp(comm, "containerd", 10) == 0 && str(((struct path *)arg0)->dentry->d_name.name) == "partitions" /
{
    @open_count[comm, str(((struct path *)arg0)->dentry->d_name.name)] = count();
}

END {
    $end_time = nsecs;
    printf("\nTracing finished. (End Time: %d)\n", $end_time);

    $duration = ($end_time - @start_time) / 1000000000; // 转换为秒
    printf("Total duration: %d seconds\n", $duration);

    printf("\nTracing results:\n");
    print(@open_count);
}
'

// output
Tracing finished. (End Time: 1684141913)
Total duration: 9 seconds

Tracing results:
@open_count[containerd-shim, partitions]: 322

// vfs_read kstack

@last_stack[120]: (
        klist_next+180   <------- spin_lock
        class_dev_iter_next+64
        disk_seqf_start+128
        show_partition_start+44
        seq_read+168
        proc_reg_read+128
        __vfs_read+72
        vfs_read+156
        ksys_read+108
        __arm64_sys_read+36
        el0_svc_handler+176
        el0_svc+8
,

yylt avatar Aug 06 '25 07:08 yylt