Re: [PATCH] aio: ensure access to ctx->ring_pages is correctly serialised

[Gu Zheng <guz.fnst@cn.fujitsu.com>]

Hi Ben,
On 03/22/2014 02:35 AM, Benjamin LaHaise wrote:

> Hi all,
> 
> Based on the issues reported by Tang and Gu, I've come up with the an 
> alternative fix that avoids adding additional locking in the event read 
> code path.  The fix is to take the ring_lock mutex during page migration, 
> which is already used to syncronize event readers and thus does not add 
> any new locking requirements in aio_read_events_ring().  I've dropped 
> the patches from Tang and Gu as a result.  This patch is now in my 
> git://git.kvack.org/~bcrl/aio-next.git tree and will be sent to Linus 
> once a few other people chime in with their reviews of this change.  
> Please review Tang, Gu.  Thanks!

As I mentioned before:
https://lkml.org/lkml/2014/3/20/34
We can put put_aio_ring_file() at the first of the
context teardown flow (aio_free_ring). Then, page migration and ctx freeing
will have mutual execution guarded by lock_page() v.s. truncate().
So that, the additional spinlock(address_space's private_lock) used to
protect use and updates of the mapping's private_data, and the sane check
of context is needless, so does the additional percpu_ref_get/put(&ctx->users).
But the enlarge ring_lock protection region to remove the additional spin_lock
is an elegant method if we ignore the effect to reading events while migrating
page is going.

Thanks,
Gu

> 
> 		-ben