Re: [PATCH] procfs: fix mmap() for /proc/vmcore

From: Al Viro
Date: Fri May 18 2018 - 22:55:42 EST


On Fri, May 18, 2018 at 08:33:37PM -0700, Linus Torvalds wrote:
> On Fri, May 18, 2018 at 8:20 PM Linus Torvalds <
> torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> > I'd *much* rather just set FMODE_UNSIGNED_OFFSET for /proc/vmcore _only_,
> > rather than open up all proc files to issues with 4G+ offsets.
>
> Hmm. I was going to point to the s_maxbytes check in rw_verify_area() and
> ask you how that ever worked for that file, but it's not there, the
> s_maxbyte checks are only in lseek and in do_splice().
>
> So apparently we protect against llseek + read/write, but we don't protect
> against pread64/pwrite64 having offset overflows..
>
> That's crazy. That makes all the s_maxbytes protection much less effective
> than it should be. Filesystems that don't get the 64-bit case right will
> screw up pread64 and friends.
>
> Al, I'm missing something. Did we always have this gaping hole where we
> didn't actually check s_maxbytes against read/write, only
> generic_file_llseek? Apparently.

Not quite. The things like
if (unlikely(*ppos >= inode->i_sb->s_maxbytes))
return 0;
iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
protect most of the regular files (see mm/filemap.c). And for devices (which is
where the majority of crap ->read()/->write() is) it's obviously not applicable -
->s_maxbytes of *what*?