Re: [PATCH] Bluetooth: virtio_bt: fix device removal

From: "Michael S. Tsirkin" <mst@redhat.com>
To: Marcel Holtmann <marcel@holtmann.org>
Cc: linux-bluetooth <linux-bluetooth@vger.kernel.org>,
	virtualization@lists.linux-foundation.org,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Luiz Augusto von Dentz <luiz.dentz@gmail.com>,
	Johan Hedberg <johan.hedberg@gmail.com>
Subject: Re: [PATCH] Bluetooth: virtio_bt: fix device removal
Date: Thu, 25 Nov 2021 20:02:39 -0500	[thread overview]
Message-ID: <20211125195924-mutt-send-email-mst@kernel.org> (raw)
In-Reply-To: <2B9668C9-427B-4D8B-A393-AAB5E50763C5@holtmann.org>

On Thu, Nov 25, 2021 at 10:58:56PM +0100, Marcel Holtmann wrote:
> Hi Michael,
> 
> >>>>> Device removal is clearly out of virtio spec: it attempts to remove
> >>>>> unused buffers from a VQ before invoking device reset. To fix, make
> >>>>> open/close NOPs and do all cleanup/setup in probe/remove.
> >>>> 
> >>>> so the virtbt_{open,close} as NOP is not really what a driver is suppose
> >>>> to be doing. These are transport enable/disable callbacks from the BT
> >>>> Core towards the driver. It maps to a device being enabled/disabled by
> >>>> something like bluetoothd for example. So if disabled, I expect that no
> >>>> resources/queues are in use.
> >>>> 
> >>>> Maybe I misunderstand the virtio spec in that regard, but I would like
> >>>> to keep this fundamental concept of a Bluetooth driver. It does work
> >>>> with all other transports like USB, SDIO, UART etc.
> >>>> 
> >>>>> The cost here is a single skb wasted on an unused bt device - which
> >>>>> seems modest.
> >>>> 
> >>>> There should be no buffer used if the device is powered off. We also don’t
> >>>> have any USB URBs in-flight if the transport is not active.
> >>>> 
> >>>>> NB: with this fix in place driver still suffers from a race condition if
> >>>>> an interrupt triggers while device is being reset. Work on a fix for
> >>>>> that issue is in progress.
> >>>> 
> >>>> In the virtbt_close() callback we should deactivate all interrupts.
> >>>> 
> >>> 
> >>> If you want to do that then device has to be reset on close,
> >>> and fully reinitialized on open.
> >>> Can you work on a patch like that?
> >>> Given I don't have the device such a rework is probably more
> >>> than I can undertake.
> >> 
> >> so you mean move virtio_find_vqs() into virtbt_open() and del_vqs() into
> >> virtbt_close()?
> > 
> > And reset before del_vqs.
> > 
> >> Or is there are way to set up the queues without starting them?
> >> 
> >> However I am failing to understand your initial concern, we do reset()
> >> before del_vqs() in virtbt_remove(). Should we be doing something different
> >> in virtbt_close() other than virtqueue_detach_unused_buf(). Why would I
> >> keep buffers attached if they are not used.
> >> 
> > 
> > They are not used at that point but until device is reset can use them.
> > Also, if you then proceed to open without a reset, and kick,
> > device will start by processing the original buffers, crashing
> > or corrupting memory.
> 
> so the only valid usage is like this:
> 
> 	vdev->config->reset(vdev);
> 
> 	while ((.. = virtqueue_detach_unused_buf(vq))) {
> 	}
> 
> 	vdev->config->del_vqs(vdev);
> 
> If I make virtbt_{open,close} a NOP, then I keep adding an extra SKB to inbuf on
> every power cycle (ifup/ifdown).

So make sure you don't :)

> How does netdev handle this?
> 
> Regards
> 
> Marcel

For net, open adds buffers to vq. close does not free them up -
they stay in the vq until device is removed.

-- 
MST

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization