lkml.org 
[lkml]   [2012]   [Aug]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [flame^Wreview] net: netprio_cgroup: rework update socket logic
On Sun, Aug 12, 2012 at 10:55:17PM -0700, John Fastabend wrote:
> On 8/12/2012 6:53 PM, Al Viro wrote:
> > Ladies and gentlemen, who the devil had reviewed that little gem?
> >
> >commit 406a3c638ce8b17d9704052c07955490f732c2b8
> >Author: John Fastabend <john.r.fastabend@intel.com>
> >Date: Fri Jul 20 10:39:25 2012 +0000
> >
> >is a bleeding bogosity that doesn't pass even the most cursory
> >inspection. It iterates through descriptor tables of a bunch
> >of processes, doing this:
> > file = fcheck_files(files, fd);
> > if (!file)
> > continue;
> >
> > path = d_path(&file->f_path, tmp, PAGE_SIZE);
> > rv = sscanf(path, "socket:[%lu]", &s);
> > if (rv <= 0)
> > continue;
> >
> > sock = sock_from_file(file, &err);
> > if (!err)
> > sock_update_netprioidx(sock->sk, p);
> >Note the charming use of sscanf() for pattern-matching. 's' (inode
> >number of socket) is completely unused afterwards; what happens here
> >is a very badly written attempt to skip non-sockets. Why, will
> >sock_from_file() blow up on non-sockets? And isn't there some less
> >obnoxious way to check that the file is a sockfs one? Let's see:
> >struct socket *sock_from_file(struct file *file, int *err)
> >{
> > if (file->f_op == &socket_file_ops)
> > return file->private_data; /* set in sock_map_fd */
> >
> > *err = -ENOTSOCK;
> > return NULL;
> >}
> >... and the first line is exactly that - a check that we are on sockfs.
> >_Far_ less expensive one, at that, so it's not even that we are avoiding
> >a costly test. In other words, all masturbation with d_path() is absolutely
> >pointless.
> >
> >Furthermore, it's racy; had been even more so before the delayed fput series
> >went in, but even now it's not safe. fcheck_files() does *NOT* guarantee
> >that file is not getting closed right now. rcu_read_lock() prevents only
> >freeing and potential reuse of struct file we'd got; it might go all the
> >way through final fput() just as we look at it. So file->f_path is not
> >protected by anything. Worse yet, neither is struct socket itself - we
> >might be going through sock_release() at the same time, so sock->sk might
> >very well be NULL, leaving us a oops even after we dump d_path() idiocy.
> >
> >To make it even funnier, there's such thing as SCM_RIGHTS datagrams and
> >descriptor passing. In other words, it's *not* going to catch all sockets
> >that would be caught by the earlier variant.
> >
>
Yes, thank you Al, I should have reviewed this more throughly.

> OK clearly I screwed it up thanks for reviewing Al. How about this.
>
> fdt = files_fdtable(files);
> for (fd = 0; fd < fdt->max_fds; fd++) {
> struct socket *sock;
> int err = 0;
>
> sock = sockfd_lookup(fd, &err);
> if (!sock) {
> lock_sock(sock->sk);
What are you protecting with lock_sock here? The other call sites don't make
use of the socket lock. Arguagbly they could, but I don't think they need to.
As long as the fd doesn't go away we should be able to update the
sk_cgrp_prioidx safely.

Regards
Neil



\
 
 \ /
  Last update: 2012-08-13 14:03    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site