linux-kernel - Re: [RFC PATCH 0/4] iov_iter: Add composite, scatterlist and skbuff iterator types

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <Z9-oaC3lVIMQ4rUF@infradead.org>
Date: Sat, 22 Mar 2025 23:21:28 -0700
From: Christoph Hellwig <hch@...radead.org>
To: David Howells <dhowells@...hat.com>
Cc: Leon Romanovsky <leonro@...dia.com>,
	Christian Brauner <christian@...uner.io>,
	Matthew Wilcox <willy@...radead.org>,
	Chuck Lever <chuck.lever@...cle.com>,
	Steve French <smfrench@...il.com>,
	Ilya Dryomov <idryomov@...il.com>, netfs@...ts.linux.dev,
	linux-fsdevel@...r.kernel.org, linux-block@...r.kernel.org,
	linux-mm@...ck.org, linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH 0/4] iov_iter: Add composite, scatterlist and skbuff
 iterator types

This is going entirely in the wrong direction.  We don't need more iter
types but less.  The reason why we have to many is because the underlying
representation of the ranges is a mess which goes deeper than just the
iterator, because it also means we have to convert between the
underlying representations all the time.

E.g. the socket code should have (and either has for a while or at least
there were patches) been using bio_vecs instead of reinventing them as sk
fragment.  The crypto code should not be using scatterlists, which are a
horrible data structure because they mix up the physical memory
description and the dma mapping information which isn't even used for
most uses, etc.

So instead of more iters let's convert everyone to a common
scatter/gather memory definition, which simplifies the iters.  For now
that is the bio_vec, which really should be converted from storing a
struct page to a phys_addr_t (and maybe renamed if that helps adoption).
That allows to trivially kill the kvec for example.

As for the head/tail - that seems to be a odd NFS/sunrpc fetish.  I've
actually started a little project to just convert the sunrpc code to
use bio_vecs, which massively simplifies the code, and allows directly
passing it to the iters in the socket API.  It doesn't quite work yet
but shows how all these custom (and in this case rather ad-hoc) memory
fragment representation cause a huge mess.

I don't think the iterlist can work in practice, but it would be nice
to have for a few use cases.  If it worked it should hopefully allow
to kill off the odd xarray iterator.