On Tue, Dec 12, 2017 at 03:13:58PM +0100, Christoph Hellwig wrote:
> On Tue, Dec 12, 2017 at 07:02:32PM +0800, Ming Lei wrote:
> > It may cause race by setting 'nvmeq' in nvme_init_request()
> > because .init_request is called inside switching io scheduler, which
> > may happen when the NVMe device is being resetted and its nvme queues
> > are being freed and created. We don't have any sync between the two
> > pathes.
> >
> > This patch removes the .init_request callback and sets the nvmeq runtime,
> > and fixes the following bug:
>
> If ->init_request doesn't work for NVMe it won't work for any other
> driver either, so we need to remove it entirely if we can't fix it.
>
> I'd much rather try to fix it first, though.
Just checked all this uses, and it is a NVMe specific issue, because only
NVMe's .init_request() associates reference of one variable to the request,
and the variable itself may change during the request's lifetime.
And all the following .init_request() have this issue:
nvme_fc_init_request()
nvme_rdma_init_request()
nvme_loop_init_request()
Thanks,
Ming
