qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH 0/6] RfC: try improve native hotplug for pcie root ports


From: Michael S. Tsirkin
Subject: Re: [PATCH 0/6] RfC: try improve native hotplug for pcie root ports
Date: Mon, 1 Nov 2021 17:47:05 -0400

On Tue, Oct 19, 2021 at 08:29:19AM +0200, Gerd Hoffmann wrote:
>   Hi,
> 
> > > Yes.  Maybe ask rh qe to run the patch set through their hotplug test
> > > suite (to avoid a apci-hotplug style disaster where qe found various
> > > issues after release)?
> > 
> > I'll poke around to see if they can help us... we'll need
> > a backport for that though.
> 
> Easy, it's a clean cherry-pick for 6.1, scratch build is on the way.
> 
> > > > I would also like to see a shorter timeout, maybe 100ms, so
> > > > that we are more responsive to guest changes in resending request.
> > > 
> > > I don't think it is a good idea to go for a shorter timeout given that
> > > the 5 seconds are in the specs and we want avoid a resent request being
> > > interpreted as cancel.
> > > It also wouldn't change anything at least for linux guests because linux
> > > is waiting those 5 seconds (with power indicator in blinking state).
> > > Only the reason for refusing 'device_del' changes from "5 secs not over
> > > yet" to "guest is busy processing the hotplug request".
> > 
> > First 5 seconds yes. But the retries afterwards?
> 
> Hmm, maybe, but I'd tend to keep it simple and go for 5 secs no matter
> what.  If the guest isn't responding (maybe because it is in the middle
> of a reboot) it's unlikely that fast re-requests are fundamentally
> changing things.
> 
> > > We could consider to tackle the 5sec timeout on the guest side, i.e.
> > > have linux skip the 5sec wait in case the root port is virtual (should
> > > be easy to figure by checking the pci id).
> > > 
> > > take care,
> > >   Gerd
> > 
> > Yes ... do we want to control how long it blinks from hypervisor side?
> 
> Is there a good reason for that?
> If not, then no.  Keep it simple.
> 
> When the guest powers off the slot pcie_cap_slot_write_config() will
> happily unplug the device without additional checks (no check whenever
> the 5 seconds are over, also no check whenever there is a pending unplug
> request in the first place).
> 
> So in theory the guest turning off slot power quickly should work just
> fine and speed up the unplug process in the common case (guest is
> up'n'running and responsitive).  Down to 1-2 secs instead of 5-7.
> Didn't actually test that though.
> 
> take care,
>   Gerd

Even if this speeds up unplug, hotplug remains slow, right?
-- 
MST




reply via email to

[Prev in Thread] Current Thread [Next in Thread]