Re: [PATCHv3 1/2] block: accumulate segment page gaps per bio

[Christoph Hellwig <hch@lst.de>]

On Tue, Aug 26, 2025 at 04:33:15PM -0600, Keith Busch wrote:
> On Tue, Aug 26, 2025 at 03:57:34PM +0200, Christoph Hellwig wrote:
> > On Tue, Aug 26, 2025 at 07:47:46AM -0600, Keith Busch wrote:
> > > Currently, the virtual boundary is always compared to bv_offset, which
> > > is a page offset. If the virtual boundary is larger than a page, then we
> > > need something like "page_to_phys(bv.bv_page) + bv.bv_offset" every
> > > place we need to check against the virt boundary.
> > 
> > bv_offset is only guaranteed to be a page offset if your use
> > bio_for_each_segment(_all) or the low-level helpers implementing
> > it and not bio_for_each_bvec(_all) where it can be much larger
> > than PAGE_SIZE.
> 
> Yes, good point. So we'd have a folio offset when it's not a single
> page, but I don't think we want to special case large folios for every
> virt boundary check. It's looking like replace bvec's "page + offset"
> with phys addrs, yeah?!

Basically everything should be using physical address.  The page + offset
is just a weird and inefficient way to represent that and we really
need to get rid of it.