On Wed, Jul 22, 2009 at 09:37:59AM -0400, Jeff Layton wrote:
> On Wed, 2009-07-22 at 14:59 +0200, Johannes Weiner wrote:
> > On Wed, Jul 22, 2009 at 07:28:19AM -0400, Jeff Layton wrote:
> > > If do_sendfile is called with a "max" value of 0, it grabs the lesser
> > > s_maxbytes value of the two superblocks to use instead. There's a
> > > problem here however. s_maxbytes is an unsigned long long and it gets
> > > cast to a signed value. If both s_maxbytes values are large enough, max
> > > will end up being negative and the comparisons in this code won't work
> > > correctly.
> > >
> > > Change do_sendfile to use unsigned values internally for the offset
> > > checks.
> > >
> > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > > ---
> > > fs/read_write.c | 6 +++---
> > > 1 files changed, 3 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/fs/read_write.c b/fs/read_write.c
> > > index 6c8c55d..36899ff 100644
> > > --- a/fs/read_write.c
> > > +++ b/fs/read_write.c
> > > @@ -788,11 +788,11 @@ SYSCALL_DEFINE5(pwritev, unsigned long, fd, const struct iovec __user *, vec,
> > > }
> > >
> > > static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
> > > - size_t count, loff_t max)
> > > + size_t count, unsigned long long max)
> > > {
> > > struct file * in_file, * out_file;
> > > struct inode * in_inode, * out_inode;
> > > - loff_t pos;
> > > + unsigned long long pos;
> > > ssize_t retval;
> > > int fput_needed_in, fput_needed_out, fl;
> > >
> > > @@ -838,7 +838,7 @@ static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
> > > if (!max)
> > > max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes);
> > >
> > > - pos = *ppos;
> > > + pos = (unsigned long long) *ppos;
> > > retval = -EINVAL;
> > > if (unlikely(pos < 0))
> > > goto fput_out;
> >
> > May it be possible that cifs is the only fs that sets sb->sb_maxbytes
> > to exceed loff_t? It seems the others clamp it to MAX_LFS_FILESIZE
> > while CIFS exceeds this by one. And then max is -1.
> >
> > So, isn't the correct fix something similar to this?
> >
> > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
> > index e16d759..df56093 100644
> > --- a/fs/cifs/connect.c
> > +++ b/fs/cifs/connect.c
> > @@ -2452,10 +2452,10 @@ try_mount_again:
> > tcon->local_lease = volume_info->local_lease;
> > }
> > if (pSesInfo) {
> > - if (pSesInfo->capabilities & CAP_LARGE_FILES) {
> > - sb->s_maxbytes = (u64) 1 << 63;
> > - } else
> > - sb->s_maxbytes = (u64) 1 << 31; /* 2 GB */
> > + if (pSesInfo->capabilities & CAP_LARGE_FILES)
> > + sb->s_maxbytes = MAX_LFS_FILESIZE;
> > + else
> > + sb->s_maxbytes = MAX_NON_LFS;
> > }
> >
> > /* BB FIXME fix time_gran to be larger for LANMAN sessions */
>
> Yes and I posted that exact same cifs patch yesterday.
Sorry, I missed it.
> I think we also need to do a similar fix for get_sb_pseudo. It
> currently sets s_maxbytes to ~0ULL...
Yes, I saw it and agree with it.
> Any of these patches will fix the immediate problem, but I think this
> code in do_sendfile should still account for the possibility that
> someone can set the value larger than MAX_LFS_FILESIZE. An alternative
> is to consider a WARN at mount time when filesystems set s_maxbytes
> larger than that value (that might help catch out of tree filesystems
> that get this wrong and prevent this sort of silent bug in the future).
Isn't MAX_LFS_FILESIZE by definition the maximum sensible value for
s_maxbytes?
> Either way, the patch I posted for this isn't sufficient since there are
> some checks that need to be done against the signed values (the
> (pos < 0) check, for instance). I'll post a respun patch in a bit that
> should fix up those problems.
That is already handled in rw_verify_area(), I think, so we should be
able to drop it completely.
Hannes
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
