On Fri, May 24, 2019 at 1:16 PM Xuehan Xu <xxhdx1985126@xxxxxxxxx> wrote:
>
> >
> > Hi Xuehan,
> >
> > While I understand the desire to use the cgroup interface to allow for
> > easy adjustment and process granularity, I think this is unlikely to be
> > accepted in the form of a new controller. Each controller is supposed
> > to distribute a specific resource and meta iops, data iops and data
> > band(width?) mostly fall under the realm of the existing I/O
> > controller. Have you run this by the cgroup folks?
> >
> > Regardless, take a look at Documentation/process/coding-style.rst for
> > rules on indentation, line length, etc. Also, the data throttle should
> > apply to rbd too, so I would change the name to "ceph".
> >
> > Thanks,
> >
> > Ilya
>
> Hi, Ilya, thanks for your review:-)
>
> I investigated the existing blkio controller before trying to
> implement a new controller. If I understand the code of blkio
> correctly, it's mainly dedicated to limiting the block device io and
> takes effect by cooperating with the io scheduler which ceph io path
> doesn't contain. So I think maybe a new controller should be
> appropriate. After all, network file system "io" is not real I/O,
> right?
"blkio" is the legacy name. This controller has been renamed to "io"
precisely because it is supposed to be able to handle any I/O, whether
to a real block device or to an unnamed instance. Writeback is wired
through backing_dev_info, which ceph instantiates like any other
network filesystem. Grep for CGROUP_WRITEBACK and SB_I_CGROUPWB.
I don't know how many sharp edges there are or if this infrastructure
is mature enough for anything other than a simple ext4-like use case,
but I wouldn't be surprised to see Tejun and others pushing back on
a ceph-specific controller.
>
> I did submit this patch to cgroup mailling list, yesterday. But no
> response has been received. I don't quite understand the procedure
> that needs to follow to contribute to the cgroup source code. Maybe I
> didn't do it right:-(
cgroups@xxxxxxxxxxxxxxx sounds right, give it a few working days.
Thanks,
Ilya
