Re: [PATCH rev2 1/4] dmaengine: enable mutliple clients and operations

[Andrew Morton <akpm@osdl.org>]

On Mon, 31 Jul 2006 23:11:58 -0700 (PDT)
David Miller <davem@davemloft.net> wrote:

> 
> Can I ask that the known bugs in the I/O AT DMA code be fixed
> before we start adding new features to it?
> 
> Specifically, the lock_cpu_hotplug() call in net_dma_rebalance()
> is still there and being invoked with a spinlock held.  The
> spinlock is grabbed by the caller, netdev_dma_event() which
> grabs the net_dma_event_lock spinlock.
> 
> You cannot invoke lock_cpu_hotplug() while holding a spinlock
> because lock_cpu_hotplug(), as seen in kernel/cpu.c, takes
> a semaphore which can sleep.  Sleeping while holding a spinlock
> is not allowed.
> 
> This is the second time I have tried to make the Intel developers
> aware of this bug.  So please fix this problem.
> 

Please just delete the lock_cpu_hotplug()/unlock_cpu_hotplug() calls.  Any
code which runs inside preempt_disable() is automatically protected from
cpu hot-unplug.

It's not presently 100% protected against cpu hot-add, but it's good enough
for now: the setting of the flag in cpu_online_map is the last thing which
happens.  To make this 100% tight we should probably run __cpu_up() via
stop_machine_run().