Re: [PATCH 01/11] scsi/fcoe: lock online CPUs in fcoe_percpu_clean()

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 04/08/2016 03:30 PM, Sebastian Andrzej Siewior wrote:
> On 03/15/2016 09:19 AM, Christoph Hellwig wrote:
>> On Fri, Mar 11, 2016 at 05:32:15PM +0100, Sebastian Andrzej Siewior wrote:
>>> alloc_workqueue() in setup and then queue_work_on(cpu, , item)? item
>>> should be struct work_struct but all I have is a skb. Is there an easy
>>> way to get this attached?
>>
>> Good question.  There is skb->cb, but it looks like it doesn't have
>> space for an additional work_item in the fcoe case.  Maybe have
>> a per-cpu work_struct and keep all the list handling as-is for now?
> 
> Okay. Let me try this. What about the few fixes from the series (which
> apply before the rework to smbboot theads)?

okay kworker. This does not look good. I have it converted what I miss
flushing work when CPU goes down and ensuring not to queue work while
the CPU is down.

- cpu_online(x) is racy. In DOWN_PREPARE the worker is deactivated /
  stopped. However slightly later the bit from the CPU mask is removed.

- Whatever is queued and did not make it before the CPU went down seems
  to be delayed until the CPU comes back online.

- if the worker keeps running while the CPU is going down the worker
  continues running on a different CPU.

So I don't see how the former two points can be solved without keeping
track of CPUs in a CPU notifier. Getting pushed to a different CPU be
probably less of an issue if we would have a work-item and would not
need to rely on the per-CPU list.

Sebastian