On Tue, Jan 31, 2012 at 12:48:49PM -0500, David Schultz wrote: > On Tue, Jan 31, 2012, Konstantin Belousov wrote: > > On Mon, Jan 30, 2012 at 02:07:03PM -0500, David Schultz wrote: > > > On Mon, Jan 30, 2012, Kostik Belousov wrote: > > > > On Sun, Jan 29, 2012 at 05:39:04PM -0500, David Schultz wrote: > > > > > On Sun, Jan 29, 2012, Kostik Belousov wrote: > > > > > > On Sat, Jan 28, 2012 at 07:12:25PM -0500, David Schultz wrote: > > > > > > > On Sat, Jan 28, 2012, Kostik Belousov wrote: > > > > > > > > On Fri, Jan 27, 2012 at 02:42:21PM -0500, David Schultz wrote: > > > > > > > > > The correct limit on the maximum size of a single read/write > > > > > > > > > is > > > > > > > > > SSIZE_MAX, but FreeBSD uses INT_MAX. It's not safe to raise > > > > > > > > > the > > > > > > > > > limit yet, though, because of bugs in several filesystems. > > > > > > > > > For > > > > > > > > > example, FFS copies uio_resid into a local variable of type > > > > > > > > > int. > > > > > > > > > I have some old patches that fix some of these issues for FFS > > > > > > > > > and > > > > > > > > > cd9660, but surely there are more places I didn't notice. > > > > > > > > > > > > > > > > > Absolutely agree. > > > > > > > > > > > > > > > > http://people.freebsd.org/~kib/misc/uio_resid.5.patch > > > > > > > > > > > > > > Nice. You found a lot more than I've got in my tree, and you even > > > > > > > fixed the return values. There are at least a few more places to > > > > > > > fix. For instance, cd9660 and the NFS client pass uio_resid or > > > > > > > iov_len to min(), which operates on ints. (Incidentally, C11 > > > > > > > generics ought to make it possible to write type-generic min() > > > > > > > and max() functions.) > > > > > > > > > > > > Thank you, http://people.freebsd.org/~kib/misc/uio_resid.6.patch > > > > > > changed them to MIN(). > > > > > > > > > > This looks good to me. I tried to think of other places that you > > > > > might have missed, and the only one that occurred to me is the > > > > Might ? I think this is a blatant understate. > > > > > > > > > pipe code. sys_pipe.c has an `int orig_resid' and lots of bogus > > > > > casts of iov_len and uio_resid to type u_int. Some look harmless, > > > > > although it appears that writing a multiple of 2^32 bytes might > > > > > result in pipe_build_write_buffer() allocating a 0-length buffer. > > > > > > > > > > My only reservation is that raising the limit could unmask a > > > > > kernel buffer overflow if we missed something, but I guess we have > > > > > to cross that bridge some day anyway. > > > > Yes, and it is an obvious reason why I am chicken to commit this for > > > > so long time. One more place, if this is reasonable to count as 'one' > > > > place, are the cdevsw methods. devfs passes uio down to the drivers. > > > > > > That's why I'm glad I'm not committing it. :) A more conservative > > > change (also known as "kicking the can down the road") would be to > > > add a VFS flag, e.g., VFCF_LONGIO, and only set it on file systems > > > that have been thoroughly reviewed. The VFS layer could cap the size > > > at INT_MAX for file systems without the flag. > > At least I will get more mail after the commit, I hope. > > > > I disagree with the VFCF_LONGIO approach. It will cause much head-scratching > > for unsuspecting user who would try to use > 4GB transfers. > > > > What I can do, is to commit all changes except removals of the checks > > for INT_MAX. After type changes settle, I can try to gather enough > > bravery to flip the checks in HEAD, possibly with temporary sysctl > > to return to old behaviour for emergency (AKA hole). > > That sounds like a good plan to me. > > As an aside, I wonder if we could convince the clang folks to add > a warning similar to `lint -a', which complains about every long->int > narrowing conversion that doesn't have an explicit cast. Modern > languages such as Java and C# require casts for narrowing > conversions, and I'd be a lot more confident about this change if > we did the same for the FreeBSD kernel. > > > > > diff --git a/sys/kern/sys_pipe.c b/sys/kern/sys_pipe.c > > > > index 9edcb74..332ec37 100644 > > > > --- a/sys/kern/sys_pipe.c > > > > +++ b/sys/kern/sys_pipe.c > > > [...] > > > > @@ -757,14 +757,14 @@ pipe_build_write_buffer(wpipe, uio) > > > > struct pipe *wpipe; > > > > struct uio *uio; > > > > { > > > > - u_int size; > > > > + size_t size; > > > > int i; > > > > > > > > PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED); > > > > KASSERT(wpipe->pipe_state & PIPE_DIRECTW, > > > > ("Clone attempt on non-direct write > > > > pipe!")); > > > > > > > > - size = (u_int) uio->uio_iov->iov_len; > > > > + size = uio->uio_iov->iov_len; > > > > if (size > wpipe->pipe_buffer.size) > > > > size = wpipe->pipe_buffer.size; > > > > > > The transfer can't be bigger than the max pipe buffer size (64k), > > > so `size = (int)MIN(uio->uio_iov->iov_len, wpipe->pipe_buffer.size)' > > > should suffice. The same comment applies elsewhere in the file. > > > > True. If you much prefer this version, I will change the patch. But I do > > think that my changes are cleaner. > > I don't mind either way. I haven't touched anything remotely > close to that code in years.
I did the changes along the way suggested by Bruce. Also, I put the patch under the real test for UFS and new NFS client, reading/writing files of sizes multiple of INT_MAX in single transaction. This indeed revealed two more issues, one in ktrace, and second in uiomove(). The later resulted in quite spectacular kernel stack corruption, because uio_iovcnt become underflowed and iovec was iterated past end. The updated patch disables SSIZE_MAX i/o by default, and includes sysctl knob debug.iosize_max_clamp, which removes the clamp if set to zero. I consider this patch as a commit candidate. http://people.freebsd.org/~kib/misc/uio_resid.9.patch
pgp3BYMbYRJ37.pgp
Description: PGP signature
