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;
Date:   Wed, 4 May 2022 13:04:35 +0200
From:   Daniel Vetter <daniel@ffwll.ch>
To:     Javier Martinez Canillas <javierm@redhat.com>
Cc:     linux-kernel@vger.kernel.org, Maxime Ripard <maxime@cerno.tech>,
        Thomas Zimmermann <tzimmermann@suse.de>,
        Daniel Vetter <daniel.vetter@ffwll.ch>,
        Daniel Vetter <daniel@ffwll.ch>, Helge Deller <deller@gmx.de>,
        dri-devel@lists.freedesktop.org, linux-fbdev@vger.kernel.org
Subject: Re: [PATCH] Revert "fbdev: Make fb_release() return -ENODEV if fbdev
 was unregistered"
Message-ID: <YnJdw89q3vxPGHQJ@phenom.ffwll.local>
Mail-Followup-To: Javier Martinez Canillas <javierm@redhat.com>,
        linux-kernel@vger.kernel.org, Maxime Ripard <maxime@cerno.tech>,
        Thomas Zimmermann <tzimmermann@suse.de>,
        Helge Deller <deller@gmx.de>, dri-devel@lists.freedesktop.org,
        linux-fbdev@vger.kernel.org
References: <20220504105140.746344-1-javierm@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220504105140.746344-1-javierm@redhat.com>
Precedence: bulk

On Wed, May 04, 2022 at 12:51:40PM +0200, Javier Martinez Canillas wrote:
> This reverts commit aafa025c76dcc7d1a8c8f0bdefcbe4eb480b2f6a. That commit
> attempted to fix a NULL pointer dereference, caused by the struct fb_info
> associated with a framebuffer device to not longer be valid when the file
> descriptor was closed.
> 
> But the solution was wrong since it was just papering over the issue, and
> also would leak any resources that might be reference counted in fb_open.
> 
> Instead, the fbdev drivers that are releasing the fb_info too soon should
> be fixed to prevent this situation to happen.

Maybe add a bit more detail here why this goes boom:

The issue was uncovered by 27599aacbaef ("fbdev: Hot-unplug firmware fb
devices on forced removal"), which added an new path that goes through the
struct device removal instead of directly unregistering the fb. Most fbdev
drivers have issues with the fb_info lifetime, because they directly call
framebuffer_release() from their device driver's ->remove callback,
instead of from fbops->fb_destroy callback. This meant that due to this
switch the fb_info was now destroyed too early, while references still
existed, while before it was simply leaked. The patch we're reverting here
reinstated that leak, hence "fixed" the regression.

With that or similar added:

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Also thinking about this, maybe we should have a WARN_ON in
framebuffer_release if the refcount is elevated, and simply bail out in
that case? That would make this bug a lot easier to debug and less
confusing.

Maybe include that in your series to fix this properly.
-Daniel

> 
> Suggested-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> Signed-off-by: Javier Martinez Canillas <javierm@redhat.com>
> ---
> 
>  drivers/video/fbdev/core/fbmem.c | 5 +----
>  1 file changed, 1 insertion(+), 4 deletions(-)
> 
> diff --git a/drivers/video/fbdev/core/fbmem.c b/drivers/video/fbdev/core/fbmem.c
> index 97eb0dee411c..a6bb0e438216 100644
> --- a/drivers/video/fbdev/core/fbmem.c
> +++ b/drivers/video/fbdev/core/fbmem.c
> @@ -1434,10 +1434,7 @@ fb_release(struct inode *inode, struct file *file)
>  __acquires(&info->lock)
>  __releases(&info->lock)
>  {
> -	struct fb_info * const info = file_fb_info(file);
> -
> -	if (!info)
> -		return -ENODEV;
> +	struct fb_info * const info = file->private_data;
>  
>  	lock_fb_info(info);
>  	if (info->fbops->fb_release)
> -- 
> 2.35.1
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch