Re: [RFC PATCH 1/4] splice: Fix corruption of spliced data after splice() returns

[Matthew Wilcox]

On Wed, Jul 19, 2023 at 09:35:33PM +0200, Miklos Szeredi wrote:
> On Wed, 19 Jul 2023 at 19:59, Matt Whitlock <kernel@xxxxxxxxxxxxxxxxx> wrote:
> >
> > On Wednesday, 19 July 2023 06:17:51 EDT, Miklos Szeredi wrote:
> > > On Thu, 29 Jun 2023 at 17:56, David Howells <dhowells@xxxxxxxxxx> wrote:
> > >>
> > >> Splicing data from, say, a file into a pipe currently leaves the source
> > >> pages in the pipe after splice() returns - but this means that those pages
> > >> can be subsequently modified by shared-writable mmap(), write(),
> > >> fallocate(), etc. before they're consumed.
> > >
> > > What is this trying to fix?   The above behavior is well known, so
> > > it's not likely to be a problem.
> >
> > Respectfully, it's not well-known, as it's not documented. If the splice(2)
> > man page had mentioned that pages can be mutated after they're already
> > ostensibly at rest in the output pipe buffer, then my nightly backups
> > wouldn't have been incurring corruption silently for many months.
> 
> splice(2):
> 
>        Though we talk of copying, actual copies are generally avoided.
> The kernel does this by implementing a pipe buffer as a set  of
> refer‐
>        ence-counted  pointers  to  pages  of kernel memory.  The
> kernel creates "copies" of pages in a buffer by creating new pointers
> (for the
>        output buffer) referring to the pages, and increasing the
> reference counts for the pages: only pointers are copied, not the
> pages of the
>        buffer.
> 
> While not explicitly stating that the contents of the pages can change
> after being spliced, this can easily be inferred from the above
> semantics.

So what's the API that provides the semantics of _copying_?  And, frankly,
this is a "you're holding it wrong" kind of argument.  It only makes
sense if you're read the implementation, which is at best level 2:

https://ozlabs.org/~rusty/index.cgi/tech/2008-03-30.html

and worst a level -5:

https://ozlabs.org/~rusty/index.cgi/tech/2008-04-01.html