RE: [PATCH] eventfd: convert to using ->write_iter()

From: David Laight <David.Laight@ACULAB.COM>
To: 'Jens Axboe' <axboe@kernel.dk>, Matthew Wilcox <willy@infradead.org>
Cc: "viro@zeniv.linux.org.uk" <viro@zeniv.linux.org.uk>,
	"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
	LKML <linux-kernel@vger.kernel.org>
Subject: RE: [PATCH] eventfd: convert to using ->write_iter()
Date: Tue, 4 May 2021 08:07:33 +0000	[thread overview]
Message-ID: <494f31b8e37b44d1a24e28885188f16e@AcuMS.aculab.com> (raw)
In-Reply-To: <7263c088-22f5-d125-cf80-5ebbd9d110e5@kernel.dk>

From: Jens Axboe
> Sent: 03 May 2021 19:05
> 
> On 5/3/21 12:02 PM, Matthew Wilcox wrote:
> > On Mon, May 03, 2021 at 11:57:08AM -0600, Jens Axboe wrote:
> >> On 5/3/21 10:12 AM, David Laight wrote:
> >>> From: Jens Axboe
> >>>> Sent: 03 May 2021 15:58
> >>>>
> >>>> Had a report on writing to eventfd with io_uring is slower than it
> >>>> should be, and it's the usual case of if a file type doesn't support
> >>>> ->write_iter(), then io_uring cannot rely on IOCB_NOWAIT being honored
> >>>> alongside O_NONBLOCK for whether or not this is a non-blocking write
> >>>> attempt. That means io_uring will punt the operation to an io thread,
> >>>> which will slow us down unnecessarily.
> >>>>
> >>>> Convert eventfd to using fops->write_iter() instead of fops->write().
> >>>
> >>> Won't this have a measurable performance degradation on normal
> >>> code that does write(event_fd, &one, 4);
> >>
> >> If ->write_iter() or ->read_iter() is much slower than the non-iov
> >> versions, then I think we have generic issues that should be solved.
> >
> > We do!
> >
> > https://lore.kernel.org/linux-fsdevel/20210107151125.GB5270@casper.infradead.org/
> > is one thread on it.  There have been others.
> 
> But then we really must get that fixed, imho ->read() and ->write()
> should go away, and if the iter variants are 10% slower, then that should
> get fixed up.

I think there are two separate issues.
(Although I've not looked in detail into the really bad cases.)

1) I suspect some of the fs code is using entirely different paths for the
'single fragment' and 'iter' variants.

2) For trivial drivers the cost of setting up the iov_iter[] and then
iterating it becomes significant (or at least measurable).

I haven't tried to undo the morass of #defines in the iter code.
But I suspect they could be optimised for the common case of
copying an entire single-fragment to/from userspace in one call.

Not related to this code path, but I've some patches that give a
few % speedup for writev() to /dev/null.
That is all about copying the iov[] from user - it doesn't get 'iterated'.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)