Re: [PATCH 13/14] block: Allow REQ_FUA|REQ_READ

[Christoph Hellwig <hch@infradead.org>]

On Mon, Mar 17, 2025 at 11:32:21AM -0600, Keith Busch wrote:
> I'm not even sure it makes sense to suspect the controller side, but the
> host side is considered reliable? If the media is in fact perfect, and
> if non-FUA read returns data that fails the checksum, wouldn't it be
> just as likely that the transport or host side is the problem?

From the error numbers we've seen it would in fact be much more likely
to be the host or transport side.

> Every
> NVMe SSD I've developered couldn't efficiently fill the PCIe tx-buffer
> directly from media (excluding Optane; RIP), so a read with or without
> FUA would stage the data in the very same controller-side DRAM.

Exactly.  It's an writeback+invalidate+reread operation.  A great
way to slow down the read, but it's not going to help you with
data recovery.