Re: [PATCH v2 6/7] md/raid1,raid10: skip futile retries on P2PDMA mapping failures

[Mykola Marzhan]

> I had thoughts about this series too. It seems best to avoid
> the P2P path entirely unless all legs support P2P.
>
> Can the md/raid code determine P2P support in advance?

Yes, and it already does: the block layer ANDs the P2PDMA queue
feature across all legs, so an array with a non-P2P member
(e.g. an nvme-tcp leg) never advertises P2P at all.

But that flag can only say "this driver can do P2P in general".
md cannot see which HCA device does the DMA for a fabrics leg —
and with multipath failover it can change at any time.

So the up-front gate covers what is knowable up front.
This series is about the remainder.
In mixed case, one local P2P leg plus a remote replica, with it:
1) a failed read just redirects to a capable mirror;
2) a failed write records bad blocks instead of retrying forever.

Mykola