Re: [PATCH] bcache: make stripe_size configurable and persistent for hardware raid5/6

[Coly Li]

On 1/6/22 11:29 AM, Eric Wheeler wrote:
> On Tue, 25 Jun 2019, Coly Li wrote:
> > On 2019/6/25 2:14 上午, Eric Wheeler wrote:
> > > On Mon, 24 Jun 2019, Coly Li wrote:
> > >
> > > > On 2019/6/23 7:16 上午, Eric Wheeler wrote:
> > > > > From: Eric Wheeler <git@xxxxxxxxxxxxxxxxxx>
> > > > >
> > > > > While some drivers set queue_limits.io_opt (e.g., md raid5), there are
> > > > > currently no SCSI/RAID controller drivers that do.  Previously stripe_size
> > > > > and partial_stripes_expensive were read-only values and could not be
> > > > > tuned by users (eg, for hardware RAID5/6).
> > > > >
> > > > > This patch enables users to save the optimal IO size via sysfs through
> > > > > the backing device attributes stripe_size and partial_stripes_expensive
> > > > > into the bcache superblock.
> > > > >
> > > > > Superblock changes are backwards-compatable:
> > > > >
> > > > > *  partial_stripes_expensive: One bit was used in the superblock flags field
> > > > >
> > > > > *  stripe_size: There are eight 64-bit "pad" fields for future use in
> > > > >     the superblock which default to 0; from those, 32-bits are now used
> > > > >     to save the stripe_size and load at device registration time.
> > > > >
> > > > > Signed-off-by: Eric Wheeler <bcache@xxxxxxxxxxxxxxxxxx>
> > > > Hi Eric,
> > > >
> > > > In general I am OK with this patch. Since Peter comments lots of SCSI
> > > > RAID devices reports a stripe width, could you please list the hardware
> > > > raid devices which don't list stripe size ? Then we can make decision
> > > > whether it is necessary to have such option enabled.
> > > Perhaps they do not set stripe_width using io_opt? I did a grep to see if
> > > any of them did, but I didn't see them. How is stripe_width indicated by
> > > RAID controllers?
> > >
> > > If they do set io_opt, then at least my Areca 1883 does not set io_opt as
> > > of 4.19.x. I also have a LSI MegaRAID 3108 which does not report io_opt as
> > > of 4.1.x, but that is an older kernel so maybe support has been added
> > > since then.
> > >
> > > Martin,
> > >
> > > Where would stripe_width be configured in the SCSI drivers? Is it visible
> > > through sysfs or debugfs so I can check my hardware support without
> > > hacking debugging the kernel?
> > >
> > > > Another point is, this patch changes struct cache_sb, it is no problem
> > > > to change on-disk format. I plan to update the super block version soon,
> > > > to store more configuration persistently into super block. stripe_size
> > > > can be added to cache_sb with other on-disk changes.
> > Hi Eric,
> >
> > > Maybe bumping version makes sense, but even if you do not, this is safe to
> > > use on systems without bumping the version because the values are unused
> > > and default to 0.
> > Yes, I understand you, it works as you suggested. I need to think how to
> > organize all options in struct cache_sb, stripe_size will be arranged
> > then. And I will ask help to you for reviewing the changes of on-disk
> > format.
> Hi Coli,
>
> Just checking in, its been a while and I didn't see any more discussion on
> the topic:

Hi Eric,

Thank you for reminding me. The persistent on-disk options were that 
much as I thought, so using a reserved space from the on-disk super 
block is fine.

> This would benefit users with older RAID controllers using RAID-5/6 that
> don't set io_opt.
>
> Even new new RAID controlers that _do_ provide `io_opt` still do _not_
> indicate partial_stripes_expensive (which is an mdraid feature, but Martin
> please correct me if I'm wrong here).  Thus, all hardware RAID-5/6 users
> could benefit by manually flagging partial_stripes_expensive to get burst
> writes out of bcache that fit their stride width.

Yeah, I agree with you.

> This patch probably needs rebased and documentation updated about io_opt,
> but here is the original patch with documentation for your reference:
> 	https://lkml.org/lkml/2019/6/22/298
>
> What do you think?

Yes please rebase the patch with latest mainline kernel and let's start 
the review.

Thank you.

Coly Li