Re: [PATCH linux-next v2] ksm: add ksm involvement information for each process

From: xu xin
Date: Sat May 11 2024 - 04:12:41 EST


>> @@ -3217,6 +3217,10 @@ static int proc_pid_ksm_stat(struct seq_file *m, struct pid_namespace *ns,
>> seq_printf(m, "ksm_zero_pages %lu\n", mm->ksm_zero_pages);
>> seq_printf(m, "ksm_merging_pages %lu\n", mm->ksm_merging_pages);
>> seq_printf(m, "ksm_process_profit %ld\n", ksm_process_profit(mm));
>> + seq_printf(m, "KSM_mergeable: %s\n",
>> + test_bit(MMF_VM_MERGEABLE, &mm->flags) ? "yes" : "no");
>
>All it *currently* means is "we called __ksm_enter()" once. It does not
>mean that KSM is still enabled for that process and that any VMA would
>be considered for merging.
>
>I don't think we should expose this.
>
>That information can be more reliably had by looking at
>
>"/proc/pid/smaps" and looking for "mg".
>
>Which tells you exactly if any VMA (and which) is currently applicable
>to KSM.
>
>
>> + seq_printf(m, "KSM_merge_any: %s\n",
>> + test_bit(MMF_VM_MERGE_ANY, &mm->flags) ? "yes" : "no");
>
>This makes more sense to export. It's the same as reading
>prctl(PR_GET_MEMORY_MERGE).
>
>The man page [1] calls it simply "KSM has been enabled for this
>process", so process-wide KSM compared to per-VMA KSM.
>
>"KSM_enabled:"
>
>*might* be more reasonable in the context of PR_SET_MEMORY_MERGE.
>
>It wouldn't tell though if KSM is enabled on the system, though.
>

I agree it. But I hope admistrators can tell if the process enabled KSM-scan
by madvise or prctl. At this point, only "/proc/pid/smaps" is not enough.

So can we add a item "KSM_enabled" which has three value as follows?

1) "prctl": KSM has been fully enabled for this process.

2) "madvise": KSM has been enabled on parts of VMA for this process.

3) "never": KSM has been never enabled for this process.

Just refer to the semantics of '/sys/kernel/mm/transparent_hugepage/enabled'