Re: [PATCH 02/15] nvme: Enforce extended LBA format for fabrics metadata

[Max Gurtovoy <maxg@mellanox.com>]

On 1/17/2020 1:53 AM, James Smart wrote:
>
>
> On 1/6/2020 5:37 AM, Max Gurtovoy wrote:
>> An extended LBA is a larger LBA that is created when metadata associated
>> with the LBA is transferred contiguously with the LBA data (AKA
>> interleaved). The metadata may be either transferred as part of the LBA
>> (creating an extended LBA) or it may be transferred as a separate
>> contiguous buffer of data. According to the NVMeoF spec, a fabrics ctrl
>> supports only an Extended LBA format. Fail revalidation in case we 
>> have a
>> spec violation. Also initialize the integrity profile for the block 
>> device
>> for fabrics ctrl.
>>
>> Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
>> Signed-off-by: Israel Rukshin <israelr@mellanox.com>
>> ---
>>   drivers/nvme/host/core.c | 25 +++++++++++++++++++++----
>>   1 file changed, 21 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
>> index d98eb48..089cdc3c 100644
>> --- a/drivers/nvme/host/core.c
>> +++ b/drivers/nvme/host/core.c
>> @@ -1818,7 +1818,7 @@ static void nvme_update_disk_info(struct 
>> gendisk *disk,
>>       blk_mq_unfreeze_queue(disk->queue);
>>   }
>>   -static void __nvme_revalidate_disk(struct gendisk *disk, struct 
>> nvme_id_ns *id)
>> +static int __nvme_revalidate_disk(struct gendisk *disk, struct 
>> nvme_id_ns *id)
>>   {
>>       struct nvme_ns *ns = disk->private_data;
>>   @@ -1846,11 +1846,21 @@ static void __nvme_revalidate_disk(struct 
>> gendisk *disk, struct nvme_id_ns *id)
>>               ns->features |= NVME_NS_EXT_LBAS;
>>             /*
>> +         * For Fabrics, only metadata as part of extended data LBA is
>> +         * supported. Fail in case of a spec violation.
>> +         */
>> +        if (ns->ctrl->ops->flags & NVME_F_FABRICS) {
>> +            if (WARN_ON_ONCE(!(ns->features & NVME_NS_EXT_LBAS)))
>> +                return -EINVAL;
>> +        }
>> +
>> +        /*
>>            * For PCI, Extended logical block will be generated by the
>>            * controller.
>>            */
>>           if (ns->ctrl->ops->flags & NVME_F_METADATA_SUPPORTED) {
>> -            if (!(ns->features & NVME_NS_EXT_LBAS))
>> +            if (ns->ctrl->ops->flags & NVME_F_FABRICS ||
>> +                !(ns->features & NVME_NS_EXT_LBAS))
>>                   ns->features |= NVME_NS_DIX_SUPPORTED;
>
> This last change seems odd - why is DIX set if NVME_F_FABRICS ?
>
> Per patch description above, Fabrics spec requires metadata as an 
> extended LBA, thus it doesn't support DIX.

we refer DIX as memory domain metadata.

>
> Which is touches on a lot of odd things with the nvme spec as it's 
> certainly possible for, within the os host implementation, to have the 
> host transmitting engine to convert an OS separate DIF buf to an 
> extended lba transmission on the wire and as presented to the 
> controller.  Transports can certainly help make this happen - and add 
> egress checking as the data leaves the host.    Which means - I'm not 
> sure this hard DIX definition being implemented this way is the way to 
> go.

RDMA transport is converting separated SGLs (non-extended mode) that 
sent by the block layer to extended mode.

The idea here is to define on which conditions we'll ask the block layer 
to set it's metadata infrastructure.

for PCI - only in case of non-extended mode (in extended mode the block 
layer will not set integrity, and the nvme driver will set the 
PRACT/PRCHK if needed) since there is no conversion to extended mode in 
the nvme driver.

for fabrics - always ask for blk integrity setting since the transport 
(RDMA only for now) is responsible for transferring it to extended mode 
on the wire.

-Max.


>
> -- james
>
>

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme