Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Fri, 31 May 2019 11:56:16 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Roman Penyaev <rpenyaev@suse.de>
Cc:     azat@libevent.org, akpm@linux-foundation.org,
        viro@zeniv.linux.org.uk, torvalds@linux-foundation.org,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v3 07/13] epoll: call ep_add_event_to_uring() from
 ep_poll_callback()
Message-ID: <20190531095616.GD17637@hirez.programming.kicks-ass.net>
References: <20190516085810.31077-1-rpenyaev@suse.de>
 <20190516085810.31077-8-rpenyaev@suse.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190516085810.31077-8-rpenyaev@suse.de>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, May 16, 2019 at 10:58:04AM +0200, Roman Penyaev wrote:
> Each ep_poll_callback() is called when fd calls wakeup() on epfd.
> So account new event in user ring.
> 
> The tricky part here is EPOLLONESHOT.  Since we are lockless we
> have to be deal with ep_poll_callbacks() called in paralle, thus
> use cmpxchg to clear public event bits and filter out concurrent
> call from another cpu.
> 
> Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Al Viro <viro@zeniv.linux.org.uk>
> Cc: Linus Torvalds <torvalds@linux-foundation.org>
> Cc: linux-fsdevel@vger.kernel.org
> Cc: linux-kernel@vger.kernel.org
> 
> diff --git a/fs/eventpoll.c b/fs/eventpoll.c
> index 2f551c005640..55612da9651e 100644
> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
> @@ -1407,6 +1407,29 @@ struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
>  }
>  #endif /* CONFIG_CHECKPOINT_RESTORE */
>  
> +/**
> + * Atomically clear public event bits and return %true if the old value has
> + * public event bits set.
> + */
> +static inline bool ep_clear_public_event_bits(struct epitem *epi)
> +{
> +	__poll_t old, flags;
> +
> +	/*
> +	 * Here we race with ourselves and with ep_modify(), which can
> +	 * change the event bits.  In order not to override events updated
> +	 * by ep_modify() we have to do cmpxchg.
> +	 */
> +
> +	old = epi->event.events;
> +	do {
> +		flags = old;
> +	} while ((old = cmpxchg(&epi->event.events, flags,
> +				flags & EP_PRIVATE_BITS)) != flags);
> +
> +	return flags & ~EP_PRIVATE_BITS;
> +}

AFAICT epi->event.events also has normal writes to it, eg. in
ep_modify(). A number of architectures cannot handle concurrent normal
writes and cmpxchg() to the same variable.