[PATCH] Add support for reservations to NVMeoF target

[Christoph Hellwig]

On Mon, Aug 28, 2017 at 09:00:20AM +0300, Sagi Grimberg wrote:
> Hi Omri,
> 
> First, please follow the format of patch submission as described
> in: Documentation/process/submitting-patches.rst
> 
> Second, I would love to see a generic reservations lib so nvmet
> and LIO can share the same abstraction (with proper callouts
> in place). But I don't think its a gating issue.
> Adding Mike to the thread.
> 
> Third, persistent reservations should survive power loss, which
> means they need to be logged in a file at least.

Strictly speaking they only need to be persistent depending on the CPTPL
field in the Reservation Register command, but implementing that is
mandatory, so yes we'll need it.  And I really want it to actually
do atomic updates, unlike the LIO code which could corrupt its database
fairly easily.

Bart (added to Cc) did some nice work on a switch and even a DLM
implementation for SCST, which I'd really like to get into mainline:

	https://github.com/bvanassche/scst/blob/master/scst/src/scst_pres.c
	https://github.com/bvanassche/scst/blob/master/scst/src/scst_dlm.c

and I'd love to get some of that code (or at least the ideas) into
mainline - both for NVMe and possibly also for the existing SCSI
target.


> As for the implementation itself, any reason to choose a rw_lock
> instead of a rcu protection scheme? Up until now we got the
> IO path to be lock free and I'm not very enthusiast starting now.

Agreed.

> > index 3b4d47a..4eb4182 100644
> > --- a/drivers/nvme/target/io-cmd.c
> > +++ b/drivers/nvme/target/io-cmd.c
> > @@ -16,6 +16,8 @@
> >   #include <linux/module.h>
> >   #include "nvmet.h"
> > +#define IEKEY	0x8	/* Ignore Existing Key */

Needs to got to include/linux/nvme.h

> > +static void nvmet_execute_resv_report(struct nvmet_req *req)
> > +{
> > +	union {
> > +		struct {
> > +			__le32	gen;
> > +			u8	rtype;
> > +			__le16	regctls;
> > +			u16	rsvd2;
> > +			u8	ptpls;
> > +			u8	rsvd14[14];
> > +		} __packed hdr;
> > +		struct {
> > +			__le16	cntlid;
> > +			u8	rcsts;
> > +			u8	rsvd5[5];
> > +			__le64	hostid;
> > +			__le64	rkey;
> > +		} __packed cntrl;
> 
> These are generic nvme structures, so they should live in
> nvme protocol header.

Agreed.   Also support for the reservation report command in nvme-cli
would be really nice for debugging.