Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <1e2134f1-f48b-1459-a38e-eac9597cd64a@kernel.dk>
Date:   Wed, 28 Jun 2023 17:14:09 -0600
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux aarch64; rv:102.0) Gecko/20100101
 Thunderbird/102.12.0
Subject: Re: [GIT PULL] bcachefs
Content-Language: en-US
To:     Kent Overstreet <kent.overstreet@linux.dev>
Cc:     torvalds@linux-foundation.org, linux-kernel@vger.kernel.org,
        linux-fsdevel@vger.kernel.org, linux-bcachefs@vger.kernel.org,
        Christoph Hellwig <hch@lst.de>,
        Christian Brauner <brauner@kernel.org>,
        Al Viro <viro@zeniv.linux.org.uk>
References: <20230627201524.ool73bps2lre2tsz@moria.home.lan>
 <c06a9e0b-8f3e-4e47-53d0-b4854a98cc44@kernel.dk>
 <20230628040114.oz46icbsjpa4egpp@moria.home.lan>
 <b02657af-5bbb-b46b-cea0-ee89f385f3c1@kernel.dk>
 <4b863e62-4406-53e4-f96a-f4d1daf098ab@kernel.dk>
 <20230628175204.oeek4nnqx7ltlqmg@moria.home.lan>
 <e1570c46-68da-22b7-5322-f34f3c2958d9@kernel.dk>
 <2e635579-37ba-ddfc-a2ab-e6c080ab4971@kernel.dk>
 <20230628221342.4j3gr3zscnsu366p@moria.home.lan>
 <d697ec27-8008-2eb6-0950-f612a602dcf5@kernel.dk>
 <20230628225514.n3xtlgmjkgapgnrd@moria.home.lan>
From:   Jens Axboe <axboe@kernel.dk>
In-Reply-To: <20230628225514.n3xtlgmjkgapgnrd@moria.home.lan>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 6/28/23 4:55?PM, Kent Overstreet wrote:
>> But it's not aio (or io_uring or whatever), it's simply the fact that
>> doing an fput() from an exiting task (for example) will end up being
>> done async. And hence waiting for task exits is NOT enough to ensure
>> that all file references have been released.
>>
>> Since there are a variety of other reasons why a mount may be pinned and
>> fail to umount, perhaps it's worth considering that changing this
>> behavior won't buy us that much. Especially since it's been around for
>> more than 10 years:
> 
> Because it seems that before io_uring the race was quite a bit harder to
> hit - I only started seeing it when things started switching over to
> io_uring. generic/388 used to pass reliably for me (pre backpointers),
> now it doesn't.

I literally just pasted a script that hits it in one second with aio. So
maybe generic/388 doesn't hit it as easily, but it's surely TRIVIAL to
hit with aio. As demonstrated. The io_uring is not hard to bring into
parity on that front, here's one I posted earlier today for 6.5:

https://lore.kernel.org/io-uring/20230628170953.952923-4-axboe@kernel.dk/

Doesn't change the fact that you can easily hit this with io_uring or
aio, and probably more things too (didn't look any further). Is it a
realistic thing outside of funky tests? Probably not really, or at least
if those guys hit it they'd probably have the work-around hack in place
in their script already.

But the fact is that it's been around for a decade. It's somehow a lot
easier to hit with bcachefs than XFS, which may just be because the
former has a bunch of workers and this may be deferring the delayed fput
work more. Just hand waving.

>> then we'd probably want to move that deferred fput list to the
>> task_struct and ensure that it gets run if the task exits rather than
>> have a global deferred list. Currently we have:
>>
>>
>> 1) If kthread or in interrupt
>> 	1a) add to global fput list
>> 2) task_work_add if not. If that fails, goto 1a.
>>
>> which would then become:
>>
>> 1) If kthread or in interrupt
>> 	1a) add to global fput list
>> 2) task_work_add if not. If that fails, we know task is existing. add to
>>    per-task defer list to be run at a convenient time before task has
>>    exited.
> 
> no, it becomes:
>  if we're running in a user task, or if we're doing an operation on
>  behalf of a user task, add to the user task's deferred list: otherwise
>  add to global deferred list.

And how would the "on behalf of a user task" work in terms of being
in_interrupt()?

-- 
Jens Axboe