Re: [PATCH 3/5] io_uring: add support for getdents

From: Dave Chinner <david@fromorbit.com>
To: "Darrick J. Wong" <djwong@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>, Hao Xu <hao.xu@linux.dev>,
	Jens Axboe <axboe@kernel.dk>,
	io-uring@vger.kernel.org,
	Dominique Martinet <asmadeus@codewreck.org>,
	Pavel Begunkov <asml.silence@gmail.com>,
	Alexander Viro <viro@zeniv.linux.org.uk>,
	Stefan Roesch <shr@fb.com>, Clay Harris <bugs@claycon.org>,
	linux-fsdevel@vger.kernel.org, Wanpeng Li <wanpengli@tencent.com>
Subject: Re: [PATCH 3/5] io_uring: add support for getdents
Date: Tue, 1 Aug 2023 08:18:27 +1000	[thread overview]
Message-ID: <ZMgzM9YF1WyJBOd/@dread.disaster.area> (raw)
In-Reply-To: <20230731152623.GC11336@frogsfrogsfrogs>

On Mon, Jul 31, 2023 at 08:26:23AM -0700, Darrick J. Wong wrote:
> On Mon, Jul 31, 2023 at 10:13:21AM +0200, Christian Brauner wrote:
> > On Mon, Jul 31, 2023 at 11:33:05AM +1000, Dave Chinner wrote:
> > > On Thu, Jul 27, 2023 at 04:27:30PM +0200, Christian Brauner wrote:
> > > But as I said in the "llseek for io-uring" thread, we need to stop
> > > the game of whack-a-mole passing random nowait boolean flags to VFS
> > > operations before it starts in earnest.  We really need a common
> > > context structure (like we have a kiocb for IO operations) that
> > > holds per operation control state so we have consistency across all
> > > the operations that we need different behaviours for.
> > 
> > Yes, I tend to agree and thought about the same. But right now we don't
> > have a lot of context. So I would lean towards a flag argument at most.
> > 
> > But I also wouldn't consider it necessarily wrong to start with booleans
> > or a flag first and in a couple of months if the need for more context
> > arises we know what kind of struct we want or need.
> 
> I'm probably missing a ton of context (because at the end of the day I
> don't care all that much about NOWAIT and still have never installed
> liburing) but AFAICT the goal seems to be that for a given io request,
> uring tries to execute it with trylocks in the originating process
> context.  If that attempt fails, it'll punt the io to a workqueue and
> rerun the request with blocking locks.  Right?

Yes, that might be the case for the VFS level code we are talking
about right now....

... but, for example, I have no clue what task context
nvmet_file_execute_rw() runs in but it definitely issues file IO
with IOCB_NOWAIT...

> I've watched quite a bit of NOWAIT whackamole going on over the past few
> years (i_rwsem, the ILOCK, the IO layer, memory allocations...).  IIRC
> these filesystem ios all have to run in process context, right?  If so,
> why don't we capture the NOWAIT state in a PF flag?  We already do that
> for NOFS/NOIO memory allocations to make sure that /all/ reclaim
> attempts cannot recurse into the fs/io stacks.

Interesting idea.

That would mean the high level code would have to toggle the task
flags before calling into the VFS, which means we'd have to capture
RWF_NOWAIT flags at the syscall level rather than propagating them
into the iocb. That may impact speed racers, because the RWF flag
propagation has been a target of significant micro-optimisation
since io_uring came along....

> "I prefer EAGAIN errors to this process blocking" doesn't seem all that
> much different.  But, what do I know...

Yeah, I can see how that might be advantageous from an API
persepective, though my gut says "that's a can of worms" but I
haven't spent enough time thinking about it to work out why I feel
that way.

-Dave.
-- 
Dave Chinner
david@fromorbit.com