Re: Lockdep warning on io_file_data_ref_zero() with 5.10-rc5

[Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>]

hi,

> On 11/28/20 5:13 PM, Pavel Begunkov wrote:
>> On 28/11/2020 23:59, Nadav Amit wrote:
>>> Hello Pavel,
>>>
>>> I got the following lockdep splat while rebasing my work on 5.10-rc5 on the
>>> kernel (based on 5.10-rc5+).
>>>
>>> I did not actually confirm that the problem is triggered without my changes,
>>> as my iouring workload requires some kernel changes (not iouring changes),
>>> yet IMHO it seems pretty clear that this is a result of your commit
>>> e297822b20e7f ("io_uring: order refnode recycling”), that acquires a lock in
>>> io_file_data_ref_zero() inside a softirq context.
>>
>> Yeah, that's true. It was already reported by syzkaller and fixed by Jens, but
>> queued for 5.11. Thanks for letting know anyway!
>>
>> https://lore.kernel.org/io-uring/948d2d3b-5f36-034d-28e6-7490343a5b59@kernel.dk/T/#t
>>
>>
>> Jens, I think it's for the best to add it for 5.10, at least so that lockdep
>> doesn't complain.
> 
> Yeah maybe, though it's "just" a lockdep issue, it can't trigger any
> deadlocks. I'd rather just keep it in 5.11 and ensure it goes to stable.
> This isn't new in this series.
Sorry, I'm not familiar with lockdep implementation, here I wonder why you say
it can't trigger any deadlocks, looking at that the syzbot report, seems that
the deadlock may happen.

And I also wonder whether spin lock bh variants are enough, normal ios are
completed in interrupt context,
==> io_complete_rw
====> __io_complete_rw
======> io_complete_rw_common
========> __io_req_complete
==========> io_put_req
============> io_free_req
==============> __io_free_req
================> io_dismantle_req
==================> io_put_file
====================> percpu_ref_put(req->fixed_file_refs);
                       if we drop the last reference here,
                       io_file_data_ref_zero() will be called,
                       then we'll call spin_lock(&data->lock);
                       in interrupt context.

Should we use spin lock irq variants?

Regards,
Xiaoguang Wang

>