In contrast, the hang reported by Mariusz Kozlowski has a slightly different feel to it, but there's a tantalizing pattern in there too:
http://www.ussg.iu.edu/hypermail/linux/kernel/0703.0/1243.html
Call Trace:
[<c03ec87e>] io_schedule+0x42/0x59
[<c0184915>] sleep_on_buffer+0x8/0xc
[<c03ed217>] __wait_on_bit+0x47/0x6c
[<c03ed297>] out_of_line_wait_on_bit+0x5b/0x64
[<c01848a8>] __wait_on_buffer+0x27/0x2d
[<c01b4228>] journal_commit_transaction+0x707/0x127f
[<c01b868b>] kjournald+0xac/0x1ed
[<c0126af5>] kthread+0xa2/0xc9
[<c010422b>] kernel_thread_helper+0x7/0x1c
which certainly also looks like an IO never completed (or completed but never woke anything up).
It also seems to be related to *buffers*. Maybe the whole bh layer thing is a fluke, but it's not waiting for normal data, it's very much waiting for those journal things that all use buffer heads.Which just makes me worry about those patches by Nick (which did come in through Andrew). I don't think it's the memorder one (it looks safe and shouldn't matter on x86 anyway!), but what about the
fs: fix __block_write_full_page error case buffer submission
locking change for example? Or that "fs: fix nobh data leak" thing with its fix? It uses "SetPageUptodate(page);" without waking up anybody who might wait for it (but the waiters here seem to wait on buffers, so that's probably not it)..
Alternatively, maybe it really is an _io_ problem (and the buffer-head thing is just a red herring, and it could happen to other IO, it's just that metadata IO uses buffer heads), and it's the scheduler changes since 2.6.20..