Re: [PATCH 3/3] blkio: Increment the blkio cgroup stats for real now

[Vivek Goyal <vgoyal@redhat.com>]

On Mon, Apr 05, 2010 at 09:53:25AM -0700, Divyesh Shah wrote:
> On Mon, Apr 5, 2010 at 8:12 AM, Vivek Goyal <vgoyal@redhat.com> wrote:
> > On Fri, Apr 02, 2010 at 04:36:34PM -0700, Divyesh Shah wrote:
> >> On Fri, Apr 2, 2010 at 12:10 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
> >> > On Thu, Apr 01, 2010 at 03:01:41PM -0700, Divyesh Shah wrote:
> >> >> We also add start_time_ns and io_start_time_ns fields to struct request
> >> >> here to record the time when a request is created and when it is
> >> >> dispatched to device. We use ns uints here as ms and jiffies are
> >> >> not very useful for non-rotational media.
> >> >>
> >> >> Signed-off-by: Divyesh Shah<dpshah@google.com>
> >> >> ---
> >> >>
> >> >>  block/blk-cgroup.c     |   60 ++++++++++++++++++++++++++++++++++++++++++++++--
> >> >>  block/blk-cgroup.h     |   14 +++++++++--
> >> >>  block/blk-core.c       |    6 +++--
> >> >>  block/cfq-iosched.c    |    4 ++-
> >> >>  include/linux/blkdev.h |   20 +++++++++++++++-
> >> >>  5 files changed, 95 insertions(+), 9 deletions(-)
> >> >>
> >> >> diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
> >> >> index ad6843f..9af7257 100644
> >> >> --- a/block/blk-cgroup.c
> >> >> +++ b/block/blk-cgroup.c
> >> >> @@ -15,6 +15,7 @@
> >> >>  #include <linux/kdev_t.h>
> >> >>  #include <linux/module.h>
> >> >>  #include <linux/err.h>
> >> >> +#include <linux/blkdev.h>
> >> >>  #include "blk-cgroup.h"
> >> >>
> >> >>  static DEFINE_SPINLOCK(blkio_list_lock);
> >> >> @@ -55,6 +56,26 @@ struct blkio_cgroup *cgroup_to_blkio_cgroup(struct cgroup *cgroup)
> >> >>  }
> >> >>  EXPORT_SYMBOL_GPL(cgroup_to_blkio_cgroup);
> >> >>
> >> >> +/*
> >> >> + * Add to the appropriate stat variable depending on the request type.
> >> >> + * This should be called with the blkg->stats_lock held.
> >> >> + */
> >> >> +void io_add_stat(uint64_t *stat, uint64_t add, unsigned int flags)
> >> >> +{
> >> >> +     if (flags & REQ_RW)
> >> >> +             stat[IO_WRITE] += add;
> >> >> +     else
> >> >> +             stat[IO_READ] += add;
> >> >> +     /*
> >> >> +      * Everywhere in the block layer, an IO is treated as sync if it is a
> >> >> +      * read or a SYNC write. We follow the same norm.
> >> >> +      */
> >> >> +     if (!(flags & REQ_RW) || flags & REQ_RW_SYNC)
> >> >> +             stat[IO_SYNC] += add;
> >> >> +     else
> >> >> +             stat[IO_ASYNC] += add;
> >> >> +}
> >> >> +
> >> >
> >> > Hi Divyesh,
> >> >
> >> > Can we have any request based information limited to cfq and not put that
> >> > in blkio-cgroup. The reason being that I am expecting that some kind of
> >> > max bw policy interface will not necessarily be implemented at CFQ
> >> > level. We might have to implement it at higher level so that it can
> >> > work with all dm/md devices. If that's the case, then it might very well
> >> > be either a bio based interface also.
> >> >
> >> > So just keeping that possibility in mind, can we keep blk-cgroup as
> >> > generic as possible and not necessarily make it dependent on "struct
> >> > request".
> >>
> >> Ok. I do understand the motivation for keeping the request related
> >> info out of blk-cgroup. Everything except the rq->cmd_flags can be
> >> easily done away with. Maybe I'll need to have CFQ send the sync and
> >> direction bits as args to the functions that need it. Not ideal coz
> >> we'll have functions with many args but I guess its not that bad too.
> >>
> >> >
> >> > If you implement, two dimensional arrays for stats then we can have
> >> > following function.
> >> >
> >> > blkio_add_stat(enum stat_type var enum stat_sub_type var_type, u64 val)
> >>
> >> I would want to avoid calls like these from CFQ into the blkcg code
> >> because many CFQ events trigger update for multiple stats (you'll see
> >> more with stats in later patchsets) and doing these calls
> >> independently for each stat would mean that we would also need to grab
> >> the stats_lock multiple times when we could've avoided that.
> >
> > I understand the need to club the updates and reduce the need of taking
> > stats_lock multiple times. I was thinking of any of following.
> >
> > - Get rid of reset interface per cgroup. Rely on changing ioscheduler on
> >  request queue and that will get rid of stats_lock entirely.
> 
> This takes away the ability to reset stats at will which is very
> useful when debugging and for testing IO controller.
> 

What do you mean by "reset stats at will"? You can change ioscheduler at
will and reset stats? The only possible issue I could think of is that only
admin can change the ioscheduler in providing per cgroup interface, one can
give write permission to indiviaul user and allow users to reset stats.

I am not sure in practice why would you allow a user to reset stats.
Especially if somebody's accounting software is based on these stats.

> > - Can we use a function blkio_add_stat() with variable number of arguments
> >  so that more than one stat can be updated in a single call?
> 
> I really don't like this at all.
> 
> > If you have other ideas to implement it without assuming "struct rq" in
> > blk-cgroup, please do that.
> 
> I've already got rid of any rq assumptions in blk-cgroup. The only
> place where we're using rq is for rq_start_time_ns() and
> rq_io-start_time_ns() functions but they are not used by the
> blk-cgroup code directly (only CFQ uses them). For another user of
> io-controller, we can implement a bio based functions.

Ok, you have made blkg->stats_lock visible to cfq. That's fine too.
Can you rename io_add_stat to blkio_add_stat. I think in V2 also, it is
still io_add_stat.

Vivek