Date:   Sun, 14 Nov 2021 17:37:15 -0500
From:   "Michael S. Tsirkin" <mst@redhat.com>
To:     Lukas Wunner <lukas@wunner.de>
Cc:     Gerd Hoffmann <kraxel@redhat.com>, linux-pci@vger.kernel.org,
        Bjorn Helgaas <bhelgaas@google.com>,
        open list <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] pciehp: fast unplug for virtual machines
Message-ID: <20211114173550-mutt-send-email-mst@kernel.org>
References: <20211111090225.946381-1-kraxel@redhat.com>
 <20211114163958.GA7211@wunner.de>
 <20211114122249-mutt-send-email-mst@kernel.org>
 <20211114180604.GA23907@wunner.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20211114180604.GA23907@wunner.de>
Precedence: bulk

On Sun, Nov 14, 2021 at 07:06:04PM +0100, Lukas Wunner wrote:
> On Sun, Nov 14, 2021 at 12:24:43PM -0500, Michael S. Tsirkin wrote:
> > On Sun, Nov 14, 2021 at 05:39:58PM +0100, Lukas Wunner wrote:
> > > Why does virtual hardware implement the Attention Button if it's
> > > perceived as annoying?  Just amend qemu so that it doesn't advertise
> > > presence of an Attention Button to get rid of the delay.  (Clear the
> > > Attention Button Present bit in the Slot Capabilities register.)
> > 
> > Because we want ability to request device removal from outside the
> > guest.
> 
> Please elaborate.  Does "outside the guest" mean on the host?
> How do you represent the Attention Button outside the guest
> and route events through to the guest?

The usual way, using kvm ioctls.

> 
> > > An Attention Button doesn't make any sense for virtual hardware
> > > except to test or debug support for it in the kernel.  Just make
> > > presence of the Attention Button optional and be done with it.
> > > 
> > > You'll still be able to bring down the slot in software via the
> > > "remove" attribute in sysfs.
> > 
> > This requires guest specific code though. Emulating the attention button
> > works in a guest independent way.
> 
> It sounds like you're using the Attention Button because it does
> almost, but not quite what you want for your specific use case.
> Now you're trying to change its behavior in a way that deviates
> from the spec to align it with your use case.
> 
> Why don't you just trigger surprise-removal from outside the guest?
> 
> Thanks,
> 
> Lukas

Because linux does not handle it well for all devices.  Fixing that
requires fixing all drivers.

-- 
MST