Re: problems with splice from /proc (was Linux 5.10-rc1)

[Christoph Hellwig]

On Tue, Oct 27, 2020 at 07:48:32AM +0100, Greg KH wrote:
> On Sun, Oct 25, 2020 at 03:40:27PM -0700, Linus Torvalds wrote:
> > The most interesting - to me - change here is Christoph's setf_fs()
> > removal (it got merged through Al Viro, as you can see in my mergelog
> > below).  It's not a _huge_ change, but it's interesting because the
> > whole model of set_fs() to specify whether a userspace copy actually
> > goes to user space or kernel space goes back to pretty much the
> > original release of Linux, and while the name is entirely historic (it
> > hasn't used the %fs segment register in a long time), the concept has
> > remained. Until now.
> 
> I told Al this yesterday, but figured I would mention it here for others
> to see.
> 
> Commit 36e2c7421f02 ("fs: don't allow splice read/write without explicit
> ops") from this patch series, is breaking the bionic test suite that
> does the following to verify that splice is working:
> 
> 	int in = open("/proc/cpuinfo", O_RDONLY);
> 	ASSERT_NE(in, -1);
> 
> 	TemporaryFile tf;
> 	ssize_t bytes_read = splice(in, nullptr, pipe_fds[1], nullptr, 8*1024, SPLICE_F_MORE | SPLICE_F_MOVE);
> 	ASSERT_NE(bytes_read, -1);
> 
> Before this change, all works well but now splice fails on /proc files
> (and I'm guessing other virtual filesystems).
> 
> I'll ask the bionic developers if they can change their test to some
> other file, but this is a regression and might show up in other "test
> platforms" as well.  Using /proc for this is just so simple because
> these files are "always there" and don't require any housekeeping for
> test suites to worry about .

Is this just a test or a real application?   I already have the
infrastructure to support read_iter/write_iter on procfs and seq_files,
but due to the intrusiveness we decided to only fix instances on an as
needed basis.  So we'll have everything ready once we pull the trigger.