On Thu, Dec 26, 2013 at 03:58:20PM +0800, Dong Aisheng wrote:

quoted

It's strange that this issue did not happen on kernel 3.10.17 with the same
code. And looking at the code, before call spin_lock we already disable the mmc
controller irq, per on my understanding, the deadlock given by lockdep may not
be able to happen(pls fix me if wrong).
May the lockdep not track the specific irq disable?
Copy lockdep guy to comment.

No, lockdep also doesn't know only that one line can take this lock.
Lockdep only knows a lock is taken from IRQ context, and disabling one
line still allows IRQs to happen and therefore it yells.