[PATCH] nvmet-passthru: Cleanup nvmet_passthru_map_sg()

[Christoph Hellwig]

On Thu, Oct 15, 2020 at 12:40:52PM -0600, Logan Gunthorpe wrote:
> > But those are entirely unrelated to the bio size.  BIO_MAX_PAGES is
> > 256, so with 4k pages and assuming none can't be merged that is 1MB,
> > while max_segments/max_hw_sectors could be something much larger.
> 
> Isn't it constrained by max_segments which is set to NVME_MAX_SEGS=127
> (for PCI)... less than BIO_MAX_PAGES...

nvmet-passthrough is not limited to work on PCIe controllers.  And even
if it did relying on an implicit limit like that is a time bomb.  I
have work pending for one of the next merge windows to lift that limit
for example, and I would not have thought that nvmet-passthrough
relies on it in any way.

> Would the NVME driver even work if max_segments was greater than
> BIO_MAX_PAGES? Correct me if I'm wrong, but it looks like
> blk_rq_map_sg() will only map one bio within a request. So there has to
> be one bio per request by the time it hits nvme_map_data()...

blk_rq_map_sg maps the bio chain.  Take a look at the for_each_bio
loop in __blk_bios_map_sg.

> If you want me to send a patch to future proof the MDTS limit with
> BIO_MAX_PAGES, I can do that, but it doesn't look like it will have any
> effect right now unless big things change.

Yes, please do.