[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] Re: [PATCH v2] expose dm-stripe target's topology I/O hints

On Thu, Aug 27 2009 at  9:12am -0400,
Alasdair G Kergon <agk redhat com> wrote:

> On Fri, Aug 21, 2009 at 10:17:18AM -0400, Mike Snitzer wrote:
> > Add .io_hints to 'struct target_type' to allow the I/O hints portion of
> > the 'struct queue_limits' to be set by each target.  Expose dm-stripe
> > target's topology I/O hints.
> Do you still think this should go into 2.6.31?

Yes, it is important.  Otherwise DM won't expose the proper I/O hints in
/sys for a striped device.  This is bad because then higher-level tools
won't be able to rely on that information to properly align ther data
(e.g. mkfs.extX -E).

> If so, please supply a revised patch header that explains the benefits
> of having the patch (and reference the new stuff in 2.6.31 that it
> uses).
> > +	limits->io_opt = chunk_size * sc->stripes;
> blk_queue_io_opt() ?

Martin didn't expose such an interface because all existing code (both
DM and MD) does not require anything beyond a simple assignment.  I
worked with Martin to establish the blk_queue_io_min() precisely because
DM needed to avoid duplicating the block layer's logic.  I think Martin
stopped short of adding a symmetric wrapper for setting io_opt because
he felt it was unnecessary.  But I could be mis-remembering.

> >  	.version = {1, 2, 0},
> Ought to inc that too when adding a new fn.

The version was already bumped during the 2.6.31-rcX cycle when I
introduced .iterate_devices (af4874e03ed82f050d5872d8c39ce64bf16b5c38).
I should bump it again?

> > +		/* Set I/O hints portion of queue limits */
> Where else are these called I/O hints or is this a new term to define
> (or avoid)?

"I/O hints" isn't formally defined in Linux code but this portion of the
I/O topology infrastructure (io_min and io_opt) is commonly referred to
as "I/O Hints".  Not a new term, it is common jargon associated with the
I/O topology support.  Martin's Linux Symposium 09 "I/O Topology" slides
refer to these as "I/O hints".  I latched on to that because it
.io_hints gives precise understanding of what portion of the
queue_limits are intended to change through this DM interface.


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]