Re: dm: set the correct discard_sectors limit

[Christoph Hellwig <hch@lst.de>]

On Sat, Mar 09, 2024 at 01:33:04PM -0500, Mike Snitzer wrote:
> That 2015 commit (0034af036554) was really ham-handed, not sure how
> I've remained unaware of this duality (with soft and hard discard
> limits) until now BUT there is quite a bit of DM code that only
> concerns itself with max_discard_sectors and discard_granularity.

Yes.

> Anyway, I'm not quite sure what you're referring to, only code that is
> still setting max_discard_sectors is drivers/md/dm.c:disable_discard

I mean all the places that I've updated to set max_discard_sectors
really should be setting max_hw_discard_sectors.  The rutime-
disabling of discard/write_zeroes/secure_discard on failure will
get new helpers one the atomic queue limits conversion is finished.

> 
> > This fixes a regression where dm bio poison v1 warns about exceeding
> > the discard bio size when running xfstests generic/500.
> 
> Meaning max_discard_sectors > max_hw_discard_sectors? What changed to
> expose this?

queue_limits_set now always recalculates max_discard_sectors from
max_hw_discard_sectors and max_user_discard_sectors.  So if a driver
sets max_discard_sectors but not max_hw_discard_sectors it will
simply be overriden.