Re: [PATCH 1/4] inet: Add skb_copy_datagram_iter

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Nov 05, 2014 at 04:57:19PM -0500, David Miller wrote:
> From: Al Viro <viro@ZenIV.linux.org.uk>
> Date: Wed, 5 Nov 2014 21:07:45 +0000
> 
> > Ping me when you put it there, OK?  I'll rebase the rest of old stuff on
> > top of it (similar helpers, mostly).
> 
> I just pushed it into net-next, thanks Al.

OK, I've taken the beginning of the old queue on top of net-next; it's
in git://git.kernel.org//pub/scm/linux/kernel/git/viro/vfs.git iov_iter-net.

>From the quick look at the remaining ->msg_iov users:

	* I'll need to add several iov_iter primitives - counterparts of
checksum.h stuff (copy_and_csum_{from,to}_iter(), maybe some more).  Not
a big deal, I'll do that tomorrow.  That will give us a clean iov_iter-based
counterpart of skb_copy_and_csum_datagram_iovec().

	* a new helper: zerocopy_sg_from_iter().  I have it, actually,
but I'd rather not step on Herbert's toes - it's too close to the areas
his series will touch, so that's probably for when his series goes in.
It will be needed for complete macvtap conversion...

	* why doesn't verify_iovec() use rw_copy_check_uvector()?  The only
real differences I see is that (a) you do allocation in callers (same as
rw_copy_check_uvector() would've done), (b) you return EMSGSIZE in case of
too long vector, while rw_copy_check_uvector() returns EINVAL in that case
and (c) you don't do access_ok().  The last one is described as optimization,
but for iov_iter primitives it's a serious PITA - for iovec-backed instances
they are using __copy_from_user()/__copy_to_user(), etc.
	It certainly would be nice to have the same code doing all copying
of iovecs from userland - readv/writev/aio/sendmsg/recvmsg/etc.  Am I
missing something subtle semantical difference in there?  EMSGSIZE vs EINVAL
is trivial (we can lift that check into the callers, if nothing else), but
I could miss something more interesting...

	* various getfrag will need to grow iov_iter-based counterparts,
but ip_append_output() needs no changes, AFAICS.

	* crypto stuff will be easy to convert - iov_iter_get_pages()
would suffice for a primitive

	* there's some really weird stuff in there.  Just what is this
static int raw_probe_proto_opt(struct flowi4 *fl4, struct msghdr *msg)
{
        struct iovec *iov;
        u8 __user *type = NULL;
        u8 __user *code = NULL;
        int probed = 0;
        unsigned int i;

        if (!msg->msg_iov)
                return 0;

        for (i = 0; i < msg->msg_iovlen; i++) {
                iov = &msg->msg_iov[i];
                if (!iov)
                        continue;
trying to do?  "If non-NULL pointer + i somehow happened to be NULL, skip it
and try to use the same pointer + i + 1"?  Huh?  Had been that way since
the function first went in back in 2004 ("[IPV4] XFRM: probe icmp type/code
when sending packets via raw socket.", according to historical tree)...

	* rds, bluetooth and vsock are doing something odd; need to RTFS some
more.

	* not sure I understand what TIPC is doing - does it prohibit too
short first segment of ->msg_iov?  net/tipc/socket.c:dest_name_check() looks
odd _and_ potentially racy - we read the same data twice and hope our checks
still apply.  I asked TIPC folks about that race back in April, but it
looks like that fell through the cracks...

Overall, so far it looks more or less feasible - other than the missing csum
primitives, current mm/iov_iter.c should suffice.  I have _not_ seriously
looked into sendpage yet; that might very well require some more.