As previously reported via Michael Reed, the FC transport took a hit
in 2.6.15 (perhaps a little earlier) when we solved a recursion error.
There are 2 deadlocks occurring:
- With scan and the delete items sharing the same workq, flushing the
workq for the delete code was getting it stalled behind a very long
running scan code path.
- There's a deadlock where scsi_remove_target() has to sit behind
scsi_scan_target() due to contention over the scan_lock().
This patch resolves the 1st deadlock and significantly reduces the
odds of the second. So far, we have only replicated the 2nd deadlock
on a highly-parallel SMP system. More on the 2nd deadlock in a following
email.
This patch reworks the transport to:
- Only use the scsi host workq for scanning
- Use 2 other workq's internally. One for deletions, the other for
scheduled deletions. Originally, we tried this with a single workq,
but the occassional flushes of the scheduled queues was hitting the
second deadlock with a slightly higher frequency. In the future, we'll
look at the LLDD's and the transport to see if we can get rid of this
extra overhead.
- When moving to the other workq's we tightened up some object states
and some lock handling.
- Properly syncs adds/deletes
- minor code cleanups
- directly reference fc_host_attrs, rather than through attribute
macros
- flush the right workq on delayed work cancel failures.
Large kudos to Michael Reed who has been working this issue for the last
month.
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
aedf349773