On 2019/3/10 下午5:29, Nikolay Borisov wrote:
>
>
> On 10.03.19 г. 5:08 ч., Anand Jain wrote:
>>
>> I agree we need btrfs specific performance measurements and its
>> my list too.
>>
>> However my idea was to add it as a btrfs-progs subcommand such as
>>
>> btrfs inspect perf ...
>>
>> And implement by using the systemtap/perf/bpf/dtrace, as these
>> can tap the kernel functions from the useland using which we
>> can measure the time taken and no kernel changes will be required.
>> But yes we need to update the btrfs-progs if we rename the kernel
>> function, which I think is ok.
>>
>> I was too early trying this with bpf before, probably there are
>> more tools now to do that same thing.
>
> This is way too developer oriented to be included in the generic btrfs
> tools. Frankly bpf makes sense but only as a separate script being
> developed and possibly shared on github or whatnot so that other
> interested people can use it. However, integrating with btrfs-progs
> definitely seems the wrong thing to do.
>
> On the same note - I'm highly against this patchset landing in the kernel.
Are you against the interface part or the btrfs specific perf part?
For the first half, indeed the new sysfs is not good and I'm OK to move
to any more established interface.
For the later half, if bfp/ftrace can do the same work, then I'm fine.
Otherwise the need is still here, both developer or experienced sys
admin will need this feature.
Thanks,
Qu
