Re: [PATCH 2/2] spapr.c: always pulse guest IRQ in spapr_core_unplug_request()

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 2845 bytes --]

On Wed, Mar 31, 2021 at 09:04:37PM -0300, Daniel Henrique Barboza wrote:
> Commit 47c8c915b162 fixed a problem where multiple spapr_drc_detach()
> requests were breaking QEMU. The solution was to just spapr_drc_detach()
> once, and use spapr_drc_unplug_requested() to filter whether we already
> detached it or not. The commit also tied the hotplug request to the
> guest in the same condition.
> 
> Turns out that there is a reliable way for a CPU hotunplug to fail. If a
> guest with one CPU hotplugs a CPU1, then offline CPU0s via 'echo 0 >
> /sys/devices/system/cpu/cpu0/online', then attempts to hotunplug CPU1,
> the kernel will refuse it because it's the last online CPU of the
> system. Given that we're pulsing the IRQ only in the first try, in a
> failed attempt, all other CPU1 hotunplug attempts will fail, regardless
> of the online state of CPU1 in the kernel, because we're simply not
> letting the guest know that we want to hotunplug the device.
> 
> Let's move spapr_hotplug_req_remove_by_index() back out of the "if
> (!spapr_drc_unplug_requested(drc))" conditional, allowing for multiple
> 'device_del' requests to the same CPU core to reach the guest, in case
> the CPU core didn't fully hotunplugged previously.
> 
> Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>

I've applied these to ppc-for-6.0, but..

> ---
>  hw/ppc/spapr.c | 11 ++++++++++-
>  1 file changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> index 05a765fab4..e4be00b732 100644
> --- a/hw/ppc/spapr.c
> +++ b/hw/ppc/spapr.c
> @@ -3777,8 +3777,17 @@ void spapr_core_unplug_request(HotplugHandler *hotplug_dev, DeviceState *dev,
>  
>      if (!spapr_drc_unplug_requested(drc)) {
>          spapr_drc_unplug_request(drc);
> -        spapr_hotplug_req_remove_by_index(drc);
>      }
> +
> +    /*
> +     * spapr_hotplug_req_remove_by_index is left unguarded, out of the
> +     * "!spapr_drc_unplug_requested" check, to allow for multiple IRQ
> +     * pulses removing the same CPU. Otherwise, in an failed hotunplug
> +     * attempt (e.g. the kernel will refuse to remove the last online
> +     * CPU), we will never attempt it again because unplug_requested
> +     * will still be 'true' in that case.
> +     */
> +    spapr_hotplug_req_remove_by_index(drc);

I think we need similar changes for all the other unplug types (LMB,
PCI, PHB) - basically retries should always be allowed, and at worst
be a no-op, rather than generating an error like they do now.

>  }
>  
>  int spapr_core_dt_populate(SpaprDrc *drc, SpaprMachineState *spapr,

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]