Re: Boot failure on Arndale with next-20131105

[Chris Mason <chris.mason@fusionio.com>]

Quoting Olof Johansson (2013-11-05 15:38:33)
> On Tue, Nov 5, 2013 at 12:33 PM, Chris Mason <chris.mason@fusionio.com> wrote:
> > Quoting Olof Johansson (2013-11-05 15:23:51)
> >> On Tue, Nov 5, 2013 at 11:33 AM, Jens Axboe <axboe@kernel.dk> wrote:
> >
> > [ horrible crashes fixed by removing my patch ]
> >
> >> > Very weird! What file system is being used?
> >>
> >> Most of my failures have happened on regular MMC cards with ext4
> >> filesystems on them.
> >>
> >> Note that the panic happens during device probe / partition table
> >> scanning, not after mounting the filesystem.
> >>
> >> Giving your patch a go now across the board. I'm very concerned about
> >> the reports of bisectability, build failures and heaps of warnings
> >> though. Did the 0-day builder pick up any of those? :-/
> >>
> >
> > Hmmm, is bcache in your config?
> 
> Doesn't look that way -- no ARM defconfigs enable it (it's what I
> build and boot), and the option defaults to off and nothing selects
> it.

Ok, I think I see it.  My guess is that you're hitting bounce buffers.
__blk_queue_bounce is the only caller of the bio splitting code I can
see that you might be hitting.

My first patch exposed a lurking bug in bio_clone_biovec.  Basically
bio->bi_vcnt is being doubled instead of initialized.

This patch is only compile tested, but I think it'll fix it.

diff --git a/fs/bio.c b/fs/bio.c
index be93de1..3595456 100644
--- a/fs/bio.c
+++ b/fs/bio.c
@@ -612,6 +612,7 @@ int bio_clone_biovec(struct bio *bio, gfp_t gfp_mask)
 	unsigned nr_iovecs = 0;
 	struct bio_vec bv, *bvl = NULL;
 	struct bvec_iter iter;
+	int i;
 
 	BUG_ON(!bio->bi_pool);
 	BUG_ON(BIO_POOL_IDX(bio) != BIO_POOL_NONE);
@@ -628,8 +629,9 @@ int bio_clone_biovec(struct bio *bio, gfp_t gfp_mask)
 		bvl = bio->bi_inline_vecs;
 	}
 
+	i = 0;
 	bio_for_each_segment(bv, bio, iter)
-		bvl[bio->bi_vcnt++] = bv;
+		bvl[i++] = bv;
 
 	bio->bi_io_vec = bvl;
 	bio->bi_iter.bi_idx = 0;