Re: splice: infinite busy loop lockup bug

From: Al Viro
Date: Mon Aug 31 2020 - 21:09:48 EST


On Mon, Aug 31, 2020 at 08:51:32PM -0400, Qian Cai wrote:
> On Fri, Aug 07, 2020 at 09:34:08PM +0900, Tetsuo Handa wrote:
> > On 2020/08/07 21:27, Al Viro wrote:
> > > On Fri, Aug 07, 2020 at 07:35:08PM +0900, Tetsuo Handa wrote:
> > >> syzbot is reporting hung task at pipe_release() [1], for for_each_bvec() from
> > >> iterate_bvec() from iterate_all_kinds() from iov_iter_alignment() from
> > >> ext4_unaligned_io() from ext4_dio_write_iter() from ext4_file_write_iter() from
> > >> call_write_iter() from do_iter_readv_writev() from do_iter_write() from
> > >> vfs_iter_write() from iter_file_splice_write() falls into infinite busy loop
> > >> with pipe->mutex held.
> > >>
> > >> The reason of falling into infinite busy loop is that iter_file_splice_write()
> > >> for some reason generates "struct bio_vec" entry with .bv_len=0 and .bv_offset=0
> > >> while for_each_bvec() cannot handle .bv_len == 0.
> > >
> > > broken in 1bdc76aea115 "iov_iter: use bvec iterator to implement iterate_bvec()",
> > > unless I'm misreading it...
>
> I have been chasing something similar for a while as in,
>
> https://lore.kernel.org/linux-fsdevel/89F418A9-EB20-48CB-9AE0-52C700E6BB74@xxxxxx/
>
> In my case, it seems the endless loop happens in iterate_iovec() instead where
> I put a debug patch here,
>
> --- a/lib/iov_iter.c
> +++ b/lib/iov_iter.c
> @@ -33,6 +33,7 @@
> if (unlikely(!__v.iov_len)) \
> continue; \
> __v.iov_base = __p->iov_base; \
> + printk_ratelimited("ITER_IOVEC left = %zu, n = %zu\n", left, n); \
> left = (STEP); \
> __v.iov_len -= left; \
> skip = __v.iov_len; \
>
> and end up seeing overflows ("n" supposes to be less than PAGE_SIZE) before the
> soft-lockups and a dead system,
>
> [ 4300.249180][T470195] ITER_IOVEC left = 0, n = 48566423
>
> Thoughts?

Er... Where does that size come from? If that's generic_perform_write(),
I'd like to see pos, offset and bytes at the time of call... ->iov_offset would
also be interesting to see (along with the entire iovec array, really).