Re: [PATCH v7 21/23] block: Avoid that flushing triggers a lockdep complaint

From: Peter Zijlstra
Date: Wed Feb 27 2019 - 09:25:02 EST


On Wed, Feb 27, 2019 at 09:35:56AM +0800, Ming Lei wrote:
> On Tue, Feb 26, 2019 at 07:08:02PM +0100, Peter Zijlstra wrote:
> > On Fri, Feb 15, 2019 at 10:26:59AM +0800, Ming Lei wrote:
> > > There might be lots of blk_flush_queue instance which is allocated
> > > for each hctx, then lots of class key slot may be wasted.
> >
> > What is 'lots' ? for someone who doesn't really know all that much about
> > the block layer.
>
> Each hw queue has one instance of blk_flush_queue, and one device may
> has lots of hw queues(may be > all possible cpus, such as nvme), and there
> may be lots of block devices in one system.
>
> Suppose one system has 10 NVMe hosts, 8 disks attached to each host, and
> 256 CPU cores in the system, there can be 10 * 8 * 256 = 20K instances of
> blk_flush_queue.
>
> Not mention there are other block devices(loop, nbd, scsi, ...) in the system.
>
> That is why I suggest to use one single lock class for addressing this
> nvme loop specific issue:
>
> https://marc.info/?l=linux-kernel&m=155019765724564&w=2

Right; that is rather a lot. But what causes the recursion, and thus how
is it specific to NVME ?