Re: [Qemu-devel] [PATCH v2 04/28] 9pfs: introduce openat_nofollow() helper

[Greg Kurz]

On Mon, 27 Feb 2017 17:28:33 -0600
Eric Blake <address@hidden> wrote:

> On 02/26/2017 04:42 PM, Greg Kurz wrote:
> > When using the passthrough security mode, symbolic links created by the
> > guest are actual symbolic links on the host file system.
> >   
> 
> > 
> > diff --git a/hw/9pfs/9p-util.c b/hw/9pfs/9p-util.c
> > new file mode 100644
> > index 000000000000..62fd7a76212a
> > --- /dev/null  
> 
> > +int openat_nofollow(int dirfd, const char *path, int flags, mode_t mode)
> > +{
> > +    int fd;
> > +
> > +    fd = dup(dirfd);
> > +    if (fd == -1) {
> > +        return -1;
> > +    }
> > +  
> 
> Do you want to assert that the caller's path does not start with '/'?

Yes, I've added this for the pull request.

> This function ignores dirfd in that case, which may not be what you want.
> 

Indeed, it really needs the path to be relative.

> > +    while (*path) {
> > +        const char *c;
> > +        int next_fd;
> > +        char *head;
> > +
> > +        head = g_strdup(path);
> > +        c = strchr(path, '/');  
> 
> So if the caller passes path="a//b", then the first iteration sets
> head="a", but the second iteration sets head="".
> 

This doesn't happen with the current code, but you're right, we should
assert here also. We only wany a/b/c/d

> 
> > +        if (c) {
> > +            head[c - path] = 0;
> > +            next_fd = openat_dir(fd, head);  
> 
> The second iteration will then fail (openat_dir on "" should fail with
> ENOENT, right?).  Oops.
> 
> > +        } else {
> > +            next_fd = openat_file(fd, head, flags, mode);
> > +        }
> > +        g_free(head);
> > +        if (next_fd == -1) {
> > +            close_preserve_errno(fd);
> > +            return -1;
> > +        }
> > +        close(fd);
> > +        fd = next_fd;
> > +
> > +        if (!c) {
> > +            break;
> > +        }
> > +        path = c + 1;  
> 
> I think the fix is that you should skip past all consecutive '/' here,
> rather than assuming there is just one.  Or can you assert that all
> callers are well-behaved, and that *path is not '/' at this point?
> 

Again you're right :-\

> > +    }  
> 
> > +static inline int openat_file(int dirfd, const char *name, int flags,
> > +                              mode_t mode)
> > +{
> > +    int fd, serrno;
> > +
> > +    fd = openat(dirfd, name, flags | O_NOFOLLOW | O_NOCTTY | O_NONBLOCK,
> > +                mode);
> > +    if (fd == -1) {
> > +        return -1;
> > +    }
> > +
> > +    serrno = errno;
> > +    /* O_NONBLOCK was only needed to open the file. Let's drop it. */
> > +    assert(!fcntl(fd, F_SETFL, flags));  
> 
> Ouch - side effect inside an assertion.  We don't support use of NDEBUG,
> but this is poor practice.
> 

And I now remember you already made a similar comment in the past... I hope
I will remember this time.

Thanks!

pgpfSPb30ABK8.pgp
Description: OpenPGP digital signature