Re: blk-mq crash under KVM in multiqueue block code (with virtio-blk and ext4)

On Thu, Sep 18, 2014 at 3:09 AM, David Hildenbrand
[off-list ref] wrote:

quoted

On Wed, Sep 17, 2014 at 10:22 PM, Jens Axboe [off-list ref] wrote:

quoted

Another way would be to ensure that the timeout handler doesn't touch hw_ctx
or tag_sets that aren't fully initialized yet. But I think this is
safer/cleaner.

That may not be easy or enough to check if hw_ctx/tag_sets are
fully initialized if you mean all requests have been used one time.

On Wed, Sep 17, 2014 at 10:11 PM, David Hildenbrand

quoted

I was playing with a simple patch that just sets cmd_flags and action_flags to

What is action_flags?

atomic_flags, sorry :)

Otherwise e.g. REQ_ATOM_STARTED could already be set due to the randomness. I
am not sure if this is really necessary, or if it is completely shielded by the
tag-handling code, but seemed to be clean for me to do it (and I remember it
not being set within blk_mq_rq_ctx_init).

You are right, both cmd_flags and atomic_flags should be cleared
in blk_mq_init_rq_map().

quoted

quoted

0. That should already be sufficient to hinder blk_mq_tag_to_rq and the calling
method to do the wrong thing.

Yes, clearing rq->cmd_flags should be enough.

And looks better to move rq initialization to __blk_mq_free_request()
too, otherwise timeout still may see old cmd_flags and rq->q before
rq's new initialization.

Yes, __blk_mq_free_request() should also reset at least rq->cmd_flags, and I
think we can remove the initialization from  __blk_mq_alloc_request().

David

quoted

Thanks,

-- 
Ming Lei