From: Al Viro
Sent: 14 April 2015 17:59
On Tue, Apr 14, 2015 at 04:36:36PM +0000, David Laight wrote:
quoted
From: Al Viro
quoted
Sent: 14 April 2015 17:34
On Tue, Apr 14, 2015 at 04:21:02PM +0000, David Laight wrote:
quoted
Massive NAK.
This breaks any code that is using msg_control to set SCTP parameters
when sending data.
Huh? ->sendmsg() expects ->msg_control already in kernel space;
it's ->recvmsg() that plays silly buggers with userland pointers there.
I read your commit message as implying that you hadn't found any
users of kernel_sendmsg() that used msg_control.
Not that the data was always read from kernel space.
Sigh... The situation is:
* ->sendmsg() expects ->msg_control copied to userland. sendmsg(2),
sendto(2), etc. do that copying. See ___sys_sendmsg() - there we have
/*
* Careful! Before this, msg_sys->msg_control contains a user pointer.
* Afterwards, it will be a kernel pointer. Thus the compiler-assisted
* checking falls down on this.
*/
if (copy_from_user(ctl_buf,
(void __user __force *)msg_sys->msg_control,
ctl_len))
goto out_freectl;
msg_sys->msg_control = ctl_buf;
As the result, ->sendmsg() instances access ->msg_control contents as normal
kernel data.
* ->recvmsg() expects ->msg_control to point to userland. See
net/core/scm.c for the helpers used to store into it. recvmsg(2) et.al.
simply leave the userland pointer there; worse, that pointer might be
to native or to compat variants, and layouts _are_ different. Thus those
if (MSG_CMSG_COMPAT & msg->msg_flags) in net/core/scm.c...
* kernel-side users of ->sendmsg() do not depend on setfs() for
access to their ->msg_control, simply because ->sendmsg() won't be using
copy_from_user()/get_user() to access it anyway.
* kernel-side users of ->recvmsg() are less lucky - most of them
don't give a damn either (they have NULL ->msg_control), but there's an
exception (somewhere in sunrpc, IIRC). So there we need to keep
playing with setfs(), even though the data side would be just fine without
that.
Apart from any other code that is using the interface.
I know you guys don't do anything to help out of tree code, but removing the setfs()
stuff from the kernel_recvmsg() code would break anything using sctp.
It shouldn't need some code lurking in sunrpc for you to leave the setfs().
In any case, how much does the setfs() cost?
I suspect it is just modifying a flag in 'current'.
A comment in kernel_recvmsg() saying that the setfs() is for msg_control
might be useful.
Then one in kelnel_sendmsg() saying that setfs() isn't needed because
msg_control is always kernel - just to avoid any confusion.
David