Re: [PATCH 00/12] unify the interface of the proportional-share policy in blkio/io

From: Paolo Valente
Date: Mon Nov 12 2018 - 10:45:48 EST




> Il giorno 12 nov 2018, alle ore 16:35, Jens Axboe <axboe@xxxxxxxxx> ha scritto:
>
> On 11/12/18 3:17 AM, Paolo Valente wrote:
>>
>>
>>> Il giorno 12 nov 2018, alle ore 11:00, Oleksandr Natalenko <oleksandr@xxxxxxxxxxxxxx> ha scritto:
>>>
>>> On 12.11.2018 10:56, Paolo Valente wrote:
>>>> Hi Jens, Tejun, all,
>>>> about nine months ago, we agreed on a solution for unifying the
>>>> interface of the proportional-share policy in blkio/io [1]. Angelo
>>>> and I finally completed it. Let me briefly recall the problem and the
>>>> solution.
>>>> The current implementation of cgroups doesn't allow two or more
>>>> entities, e.g., I/O schedulers, to share the same files. So, if CFQ
>>>> creates its files for the proportional-share policy, such as, e.g,
>>>> weight files for blkio/io groups, BFQ cannot attach somehow to them.
>>>> Thus, to enable people to set group weights with BFQ, I resorted to
>>>> making BFQ create its own version of these common files, by prepending
>>>> a bfq prefix.
>>>> Actually, no legacy code uses these different names, or is likely to
>>>> do so. Having these two sets of names is simply a source of
>>>> confusion, as pointed out also, e.g., by Lennart Poettering (CCed
>>>> here), and acknowledged by Tejun [2].
>>>> In [1] we agreed on a solution that solves this problem, by actually
>>>> making it possible to share cgroups files. Both writing to and
>>>> reading from a shared file trigger the appropriate operation for each
>>>> of the entities that share the file. In particular, in case of
>>>> reading,
>>>> - if all entities produce the same output, the this common output is
>>>> shown only once;
>>>> - if the outputs differ, then every per-entity output is shown,
>>>> preceded by the name of the entity that produced that output.
>>>> With this solution, legacy code that, e.g., sets group weights, just
>>>> works, regardless of the I/O scheduler actually implementing
>>>> proportional share.
>>>> But note that this extension is not restricted to only blkio/io. The
>>>> general group interface now enables files to be shared among multiple
>>>> entities of any kind.
>>>> (I have also added a patch to fix some clerical errors in bfq doc,
>>>> which I found while making the latter consistent with the new
>>>> interface.)
>>>> Thanks,
>>>> Paolo
>>>> [1] https://lkml.org/lkml/2018/1/4/667
>>>> [2] https://github.com/systemd/systemd/issues/7057
>>>> Angelo Ruocco (7):
>>>> kernfs: add function to find kernfs_node without increasing ref
>>>> counter
>>>> cgroup: link cftypes of the same subsystem with the same name
>>>> cgroup: add owner name to cftypes
>>>> block, bfq: align min and default weights with cfq
>>>> cgroup: make all functions of all cftypes be invoked
>>>> block, cfq: allow cgroup files to be shared
>>>> block, throttle: allow sharing cgroup statistic files
>>>> Paolo Valente (5):
>>>> cgroup: add hook seq_show_cft with also the owning cftype as parameter
>>>> block, cgroup: pass cftype to functions that need to use it
>>>> block, bfq: use standard file names for the proportional-share policy
>>>> doc, bfq-iosched: fix a few clerical errors
>>>> doc, bfq-iosched: make it consistent with the new cgroup interface
>>>> Documentation/block/bfq-iosched.txt | 31 +++--
>>>> block/bfq-cgroup.c | 148 +++++++++++++-------
>>>> block/bfq-iosched.h | 4 +-
>>>> block/blk-cgroup.c | 22 +--
>>>> block/blk-throttle.c | 24 ++--
>>>> block/cfq-iosched.c | 105 +++++++++++----
>>>> fs/kernfs/dir.c | 13 ++
>>>> include/linux/blk-cgroup.h | 10 +-
>>>> include/linux/cgroup-defs.h | 14 +-
>>>> include/linux/cgroup.h | 13 ++
>>>> include/linux/kernfs.h | 7 +
>>>> kernel/cgroup/cgroup.c | 262 +++++++++++++++++++++++++++++-------
>>>> 12 files changed, 483 insertions(+), 170 deletions(-)
>>>> --
>>>> 2.16.1
>>>
>>> I thought all the legacy stuff including CFS et al. is going to be removed in v4.21 completelyâ
>>>
>>
>> Thanks for pointing this out.
>>
>> People with a lower kernel version than the future 4.21 just cannot
>> and will not be able to use the proportional share policy on blk-mq
>> (with legacy code), because of the name issue highlighted in this
>> email. If this patch series gets accepted, a backport will solve the
>> problem. In this respect, such a backport might even happen
>> 'automatically', as most bfq commit seem to get backported to older,
>> stable kernels.
>>
>> In addition, this extension
>> - extends the whole cgroups interface, in a seamless and
>> backward-compatible way, to prevent future issues like these;
>> - solves a similar issue with throttle (which AFAIK won't go away
>> with 4.21).
>
> There's no way this series can get accepted, since you've made the
> mistake of basing it on something that won't apply to the block
> tree for 4.21.

Of course, sorry :(

We'll rebase V2.

BTW, since this patch series is probably even more useful for older
than for future kernels, might it make sense to also propose it for
stable/longterm kernels (provided that such a possibility exists)?

Thanks,
Paolo

> I've outlined these rules before, but here they are
> again:
>
> 1) Patches destined for the CURRENT kernel version should be
> against my for-linus branch. That means that right now, any
> patches that should to into 4.20 should be against that.
>
> 2) Patches destined for the NEXT kernel version should be against
> my for-x.y/block branch, where x.y is the next version. As of
> right now, patches for 4.21 should be against my for-4.21/bloc
> branch.
>
> I'd encourage you to respin against that, particularly in this case
> since we've both got a lot of churn, and also removal of various
> items that you are patching here.
>
> --
> Jens Axboe