Re: New TRIM/UNMAP tree published (2009-05-02)

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Sun, 2009-05-03 at 14:40 -0400, Jeff Garzik wrote:
> Jeff Garzik wrote:
> > (2) determine at init if queue (a) supports explicit DISCARD and/or (b) 
> > supports DISCARD flag passed with READ or WRITE
> 
> 
> As an aside -- does any existing command set support case #b, above?

Not to my knowledge.

I think discard was modelled on barrier (which can be associated with
data or stand on its own).

> AFAICT, ATA, SCSI and NVMHCI all have a single, explicit hardware 
> command to discard/deallocate unused sectors.
> 
> Therefore, creating REQ_TYPE_DISCARD seems to eliminate any need for new 
> hook ->prepare_discard().

Well, yes and no ... in the SCSI implementation of either UNMAP or WRITE
SAME, we still need a data buffer to store either the extents or the
actual same data.

> This provides a 1:1 correspondence between hardware and struct request, 
> most closely matching the setup of known hardware.

Agreed ... but we still have to allocate the adjunct buffer
somewhere ....

James