On 09/05/2012 02:39 AM, Christoph Lameter wrote:
> On Mon, 3 Sep 2012, Glauber Costa wrote:
>
>> Here is the code for that in slab_common.c:
>>
>> if (!__kmem_cache_shutdown(s)) {
>> if (s->flags & SLAB_DESTROY_BY_RCU)
>> rcu_barrier();
>>
>> __kmem_cache_destroy(s);
>> } ...
>>
>> All that code that used to belong in __kmem_cache_destroy(), will not be
>> executed in kmem_cache_shutdown() without an rcu_barrier.
>
> But that allocator specific code in __kmem_cache_destroy will not free the
> kmem_cache structure. That is the only important thing to be aware of.
> Only deferred frees of slab pages may still be in progress at this time
> until the close of the RCU period. These deferred freeing actions do not
> refer to anything but the kmem_cache structure. Therefore the rest can be
> freed before the period is over. And we check that the rest can be freed.
> Should there be a leftover at that point then f.e.
> free_partial() will issue a warning.
>
Ok. That sounds reasonable.
(not sure if correct, but reasonable)
> kmem_cache_destroy() can only be called after all objects have been freed
> and it checks that this actually was done. "Have been freed" means in the
> context of an SLAB_DESTROY_BY_RCU slab that the rcu delayed frees for the
> individual objects are complete. During kmem_cache_destroy() only slab
> pages that contain no objects are freed back to the page allocator. Those
> will be also freed in a deferred way at kmem_cache_destroy. Hmmm.... we
> could simply delete the SLAB_DESTROY_BY_RCU flag and free the objects
> without obeying the rcu period since no objects should be allocated at
> that point.
>
>> You need at least Paul's ack here to guarantee it is safe, but I believe
>> it is not. Take a look for instance at 7ed9f7e5db5, which describes a
>> subtle bug arising from such a situation.
>
> The commit that you referred to ensures that kmem_cache is not freed
> before the rcu period is over. This patch does not change that guarantee.
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxx. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>