On Wed, Feb 10, 2021 at 10:11:15PM +0000, Chaitanya Kulkarni wrote:

On 2/10/21 5:20 AM, Christoph Hellwig wrote:

quoted

MDTS vs Write Zeroes is a common misconception.  We had that discussion
in the NVMe working group, and while the text should pretty clear that
MDTS only applies to data transfers many implementators did not understand
that, which is why we added the clarification you quoted.

Also the next NVMe spec will allow devices to advertise an explicit
Write Zeroes limit, take a look at TP4040 from the "NVM Express 1.4 Ratified
TPs" at https://nvmexpress.org/developers/nvme-specification/.

So I think limiting to MDTS unless the new WZSL is set might be a fail
safe option, even if it is more pessimistic than what the spec says.

I'd also love to retest most Write Zeroes quirks with that in place.

Do you prefer some variant of the following patch (totally untested)? OR
something else ?

Somwhat.  As said I suspect defaulting to MDTS with a big fat comment
might make most sense unless the new WZSL is set.

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