On 2/27/20 12:53 AM, Christoph Hellwig wrote:

On Wed, Feb 26, 2020 at 08:58:46AM -0700, Jens Axboe wrote:

quoted

Yeah, should probably be a RWF_ flag instead, and a 64-bit SQE field
for the PI data. The 'last iovec is PI' is kind of icky.

Abusing an iovec (although I though of the first once when looking
into it) looks really horrible, but has two huge advantages:

 - it doesn't require passing another argument all the way down
   the I/O stack
 - it works with all the vectored interfaces that take a flag
   argument, so not just io_uring, but also preadv2/pwritev2 and aio.
   And while I don't care too much about the last I think preadv2
   and pwritev2 are valuable to support.

Indeed, actually the 'last iovec is PI' idea was learned from Darrick's original
patch which support PI passthrough via aio.
https://www.mail-archive.com/linux-scsi@vger.kernel.org/msg27537.html