Re: Mysterious CFQ crash and RCU

[Jens Axboe <jaxboe@fusionio.com>]

On 2011-06-05 12:38, Paul Bolle wrote:
> On Sun, 2011-06-05 at 10:39 +0200, Paul Bolle wrote:
>> 0) I'd guess not, as the last thing I tried before simply ripping
>> io_context.ioc_data out, was:
>>
>>         spin_lock_irqsave(&ioc->lock, flags);
>>         rcu_read_lock();
>>         ioc_data = rcu_dereference(ioc->ioc_data);
>>         rcu_read_unlock();
>>         if (ioc_data == cic)
>>                 rcu_assign_pointer(ioc->ioc_data, NULL);
>>         spin_unlock_irqrestore(&ioc->lock, flags);

You don't need to nest a rcu_read_lock/unlock() inside that write lock.

> 0) Strike that. It does seem to fix it. I must have done something silly
> while doing yet another test of this stuff (perhaps by forgetting to
> remove an old, crashing version of the cfq_iosched module before
> installing a new version and switching back to cfq).

Good, I thought it should!

> 1) Jens, I suggest I'll send in a small series of patches, along these
> lines:
> - cleanup: make two functions static in block/cfq-iosched.c
> - cleanup: rename io_context.cic_list -> io_context.cic_hlist
>            rename io_context.ioc_data -> io_context.last_cic
>            change io_context.last_cic from void* to cfq_io_context* 
>            rename cfq_io_context.cic_list -> cfq_io_context.cic_node
> - cleanup: rename *ret -> *ioc in two functions
> - locking: make sure io_context.last_cic is accessed under
>            io_context.lock and using the required RCU voodoo
> 
> 2) Any objections beforehand?

I've already applied the fix I sent to you, since it was definitely a
bug. So please send anything around that. Was out today, so not pushed
out until now and just now catching up on the email.

And thanks a ton for being persistent and doing a ton of work trying to
figure out what was wrong. Fixing is the easy part, illuminating where
the problem is or may be is the hard part!

-- 
Jens Axboe