On Wed, Jul 22, 2009 at 10:13:52AM -0400, Jeff Layton wrote:
> On Wed, 2009-07-22 at 15:51 +0200, Johannes Weiner wrote:
>
> > > 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?
> >
>
> Pretty much, but nothing seems to enforce it or let you know when you've
> exceeded it.
Kernel code shoots from the hips. We're faster that way. We might
shoot granny in the process.
> > > 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.
>
> If we get rid of those checks altogether, then "max" will become unused.
> Is that really OK here?
We still need to check for exceeding s_maxbytes, the other checks are
redundant.
> For discussion purposes, I've attached a replacement patch that I'm
> working with now.
> >>From 00a22f2f1e34ba0765ca49120499e681477a265a Mon Sep 17 00:00:00 2001
> From: Jeff Layton <jlayton@xxxxxxxxxx>
> Date: Wed, 22 Jul 2009 08:36:22 -0400
> Subject: [PATCH] fix offset checks in do_sendfile to use unsigned values (try #2)
>
> This is the second version of this patch. Some of the checks do need
> to use signed values. This patch should be more correct in that regard.
> This also adds a check to make sure that "pos + count" doesn't
> overflow.
>
> 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 against "max".
>
> Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> ---
> fs/read_write.c | 21 ++++++++++++---------
> 1 files changed, 12 insertions(+), 9 deletions(-)
>
> diff --git a/fs/read_write.c b/fs/read_write.c
> index 6c8c55d..2c5b402 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, newpos;
> ssize_t retval;
> int fput_needed_in, fput_needed_out, fl;
>
> @@ -835,14 +835,16 @@ static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
> goto fput_out;
> count = retval;
>
> - if (!max)
> - max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes);
> -
> - pos = *ppos;
> retval = -EINVAL;
> - if (unlikely(pos < 0))
> + if (unlikely(*ppos < 0))
> goto fput_out;
That check is done in rw_verify_area().
> - if (unlikely(pos + count > max)) {
> +
> + if (!max)
> + max = min(in_inode->i_sb->s_maxbytes,
> + out_inode->i_sb->s_maxbytes);
> + pos = (unsigned long long) *ppos;
> + newpos = pos + count;
> + if (unlikely(newpos > max || newpos < count)) {
pos + count overflow is checked in rw_verify_area() as well.
> retval = -EOVERFLOW;
> if (pos >= max)
> goto fput_out;
> @@ -869,7 +871,8 @@ static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
>
> inc_syscr(current);
> inc_syscw(current);
> - if (*ppos > max)
> + pos = (unsigned long long) *ppos;
> + if (pos > max)
> retval = -EOVERFLOW;
This one is needed.
But frankly, I really don't like the approach of catching a bogus
s_maxbytes in do_sendfile(). If we want to sanity check s_maxbytes,
we should do it at mount time.
And allowing for a bigger s_maxbytes makes no sense if the data types
involved when accessing the file can not handle these offsets at all,
no?
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
