[PATCH 06/14] block: lift setting the readahead size into the block layer

[Mike Snitzer]

On Thu, Sep 10 2020 at  5:28am -0400,
Christoph Hellwig <hch at lst.de> wrote:

> On Wed, Sep 02, 2020 at 12:20:07PM -0400, Mike Snitzer wrote:
> > On Wed, Sep 02 2020 at 11:11am -0400,
> > Christoph Hellwig <hch at lst.de> wrote:
> > 
> > > On Wed, Aug 26, 2020 at 06:07:38PM -0400, Mike Snitzer wrote:
> > > > On Sun, Jul 26 2020 at 11:03am -0400,
> > > > Christoph Hellwig <hch at lst.de> wrote:
> > > > 
> > > > > Drivers shouldn't really mess with the readahead size, as that is a VM
> > > > > concept.  Instead set it based on the optimal I/O size by lifting the
> > > > > algorithm from the md driver when registering the disk.  Also set
> > > > > bdi->io_pages there as well by applying the same scheme based on
> > > > > max_sectors.
> > > > > 
> > > > > Signed-off-by: Christoph Hellwig <hch at lst.de>
> > > > > ---
> > > > >  block/blk-settings.c         |  5 ++---
> > > > >  block/blk-sysfs.c            |  1 -
> > > > >  block/genhd.c                | 13 +++++++++++--
> > > > >  drivers/block/aoe/aoeblk.c   |  2 --
> > > > >  drivers/block/drbd/drbd_nl.c | 12 +-----------
> > > > >  drivers/md/bcache/super.c    |  4 ----
> > > > >  drivers/md/dm-table.c        |  3 ---
> > > > >  drivers/md/raid0.c           | 16 ----------------
> > > > >  drivers/md/raid10.c          | 24 +-----------------------
> > > > >  drivers/md/raid5.c           | 13 +------------
> > > > >  10 files changed, 16 insertions(+), 77 deletions(-)
> > > > 
> > > > 
> > > > In general these changes need a solid audit relative to stacking
> > > > drivers.  That is, the limits stacking methods (blk_stack_limits)
> > > > vs lower level allocation methods (__device_add_disk).
> > > > 
> > > > You optimized for lowlevel __device_add_disk establishing the bdi's
> > > > ra_pages and io_pages.  That is at the beginning of disk allocation,
> > > > well before any build up of stacking driver's queue_io_opt() -- which
> > > > was previously done in disk_stack_limits or driver specific methods
> > > > (e.g. dm_table_set_restrictions) that are called _after_ all the limits
> > > > stacking occurs.
> > > > 
> > > > By inverting the setting of the bdi's ra_pages and io_pages to be done
> > > > so early in __device_add_disk it'll break properly setting these values
> > > > for at least DM afaict.
> > > 
> > > ra_pages never got inherited by stacking drivers, check it by modifying
> > > it on an underlying device and then creating a trivial dm or md one.
> > 
> > Sure, not saying that it did.  But if the goal is to set ra_pages based
> > on io_opt then to do that correctly on stacking drivers it must be done
> > in terms of limits stacking right?  Or at least done at a location that
> > is after the limits stacking has occurred?  So should DM just open-code
> > setting ra_pages like it did for io_pages?
> > 
> > Because setting ra_pages in __device_add_disk() is way too early for DM
> > -- given it uses device_add_disk_no_queue_reg via add_disk_no_queue_reg
> > at DM device creation (before stacking all underlying devices' limits).
> 
> I'll move it to blk_register_queue, which should work just fine.

That'll work for initial DM table load as part of DM device creation
(dm_setup_md_queue).  But it won't account for DM table reloads that
might change underlying devices on a live DM device (done using
__bind).

Both dm_setup_md_queue() and __bind() call dm_table_set_restrictions()
to set/update queue_limits.  It feels like __bind() will need to call a
new block helper to set/update parts of queue_limits (e.g. ra_pages and
io_pages).

Any chance you're open to factoring out that block function as an
exported symbol for use by blk_register_queue() and code like DM's
__bind()?

Thanks,
Mike