Hello, Tahsin.

On Thu, Mar 02, 2017 at 02:33:11PM -0800, Tahsin Erdogan wrote:

quoted

And let blkg_create() verify these conditions after releasing and
regrabbing the lock.

This also means that the init path can simply pass in GFP_KERNEL.

I tried that approach, but I encountered two issues that complicate things:

1) Pushing down blk_queue_bypass(q) check in blkg_create() doesn't
quite work because when blkcg_init_queue() calls blkg_create(), the
queue is still in bypassing mode.

2) Pushing down blkcg_policy_enabled() doesn't work well either,
because blkcg_init_queue() doesn't have a policy to pass down. We
could let it pass a NULL parameter but that would make blkg_create
more ugly.

I see.  It kinda really bothers me that we'll have two different modes
for non-atomic allocations.  Can't we bind both to the policy
parameter?  Skip the checks if policy is NULL?

Thanks.

-- 
tejun