Patchwork versatile: Push lsi initialization to the end

login
register
mail settings
Submitter Paolo Bonzini
Date Oct. 8, 2012, 3:18 p.m.
Message ID <5072EEB7.2070408@redhat.com>
Download mbox | patch
Permalink /patch/190048/
State New
Headers show

Comments

Paolo Bonzini - Oct. 8, 2012, 3:18 p.m.
Il 08/10/2012 08:52, Jan Kiszka ha scritto:
> On 2012-10-06 04:13, Peter Maydell wrote:
>> On 5 October 2012 19:01, Jan Kiszka <jan.kiszka@siemens.com> wrote:
>>> I'm not a fan of this either, but the alternatives are way more
>>> complicated. We either need to rewrite the chardev subsystem,
>>> specifically how mux'ed devices are registered and how the active one is
>>> selected. Or we need to avoid flushing "unrelated" BHs for block
>>> devices. Not sure of those read requests can be postponed.
>>
>> Is this a regression? If it is then the obvious answer is to back
>> out whatever broke it...
> 
> I'm using this machine for the first time, so I cannot answer this from
> the top of my head. However, I don't think it can be a regression.
> 
> Mux chardevs work like this: You create the backend, then you register
> the frontend with them, one by one. The last one registered is the first
> one active. It should also receive the open event of chardev. But as
> that open even is issued via a BH and last frontend, the serial device,
> arrives after the first BH flushing, things break.

Does something like this work instead?

 static CharDriverState *qemu_chr_open_mux(CharDriverState *drv)

Paolo
Jan Kiszka - Oct. 8, 2012, 3:28 p.m.
On 2012-10-08 17:18, Paolo Bonzini wrote:
> Il 08/10/2012 08:52, Jan Kiszka ha scritto:
>> On 2012-10-06 04:13, Peter Maydell wrote:
>>> On 5 October 2012 19:01, Jan Kiszka <jan.kiszka@siemens.com> wrote:
>>>> I'm not a fan of this either, but the alternatives are way more
>>>> complicated. We either need to rewrite the chardev subsystem,
>>>> specifically how mux'ed devices are registered and how the active one is
>>>> selected. Or we need to avoid flushing "unrelated" BHs for block
>>>> devices. Not sure of those read requests can be postponed.
>>>
>>> Is this a regression? If it is then the obvious answer is to back
>>> out whatever broke it...
>>
>> I'm using this machine for the first time, so I cannot answer this from
>> the top of my head. However, I don't think it can be a regression.
>>
>> Mux chardevs work like this: You create the backend, then you register
>> the frontend with them, one by one. The last one registered is the first
>> one active. It should also receive the open event of chardev. But as
>> that open even is issued via a BH and last frontend, the serial device,
>> arrives after the first BH flushing, things break.
> 
> Does something like this work instead?
> 
> diff --git a/qemu-char.c b/qemu-char.c
> index b082bae..1ed6d49 100644
> --- a/qemu-char.c
> +++ b/qemu-char.c
> @@ -465,6 +465,9 @@ static void
> mux_chr_update_read_handler(CharDriverState *chr)
>      d->focus = d->mux_cnt;
>      d->mux_cnt++;
>      mux_chr_send_event(d, d->focus, CHR_EVENT_MUX_IN);
> +    if (chr->opened) {
> +        mux_chr_send_event(d, d->focus, CHR_EVENT_OPENED);
> +    }

It's not (only) about a missing event for the serial frontend, it's also
about a spurious open event to the monitor. That generates unwanted
output during startup.

Jan
Peter Maydell - Oct. 8, 2012, 4:02 p.m.
On 8 October 2012 16:28, Jan Kiszka <jan.kiszka@siemens.com> wrote:
> It's not (only) about a missing event for the serial frontend, it's also
> about a spurious open event to the monitor. That generates unwanted
> output during startup.

Right, so whatever's sending that event needs not to do so until
creation of the machine is complete. Sending events randomly
halfway through init is a recipe for problems...

-- PMM
Paolo Bonzini - Oct. 8, 2012, 4:28 p.m.
Il 08/10/2012 18:02, Peter Maydell ha scritto:
>> > It's not (only) about a missing event for the serial frontend, it's also
>> > about a spurious open event to the monitor. That generates unwanted
>> > output during startup.
> Right, so whatever's sending that event needs not to do so until
> creation of the machine is complete. Sending events randomly
> halfway through init is a recipe for problems...

So we now have proof that using bottom halves outside the block layer
was a bad idea.  Thanks Jan. :)

If Stefan goes on with the AioContext idea, we can have separate
AioContexts for qemu_bh_new on one side and for block devices on the
other, so that qemu_bh_new bottom halves won't fire just because
something is calling the block layer.

In the meanwhile, qemu_char could use a QEMUTimer expiring in the past
instead of a bottom half.

Paolo
Jan Kiszka - Oct. 8, 2012, 4:46 p.m.
On 2012-10-08 18:28, Paolo Bonzini wrote:
> Il 08/10/2012 18:02, Peter Maydell ha scritto:
>>>> It's not (only) about a missing event for the serial frontend, it's also
>>>> about a spurious open event to the monitor. That generates unwanted
>>>> output during startup.
>> Right, so whatever's sending that event needs not to do so until
>> creation of the machine is complete. Sending events randomly
>> halfway through init is a recipe for problems...
> 
> So we now have proof that using bottom halves outside the block layer
> was a bad idea.  Thanks Jan. :)
> 
> If Stefan goes on with the AioContext idea, we can have separate
> AioContexts for qemu_bh_new on one side and for block devices on the
> other, so that qemu_bh_new bottom halves won't fire just because
> something is calling the block layer.
> 
> In the meanwhile, qemu_char could use a QEMUTimer expiring in the past
> instead of a bottom half.

Sounds "beautiful", waiting to break when we leave like this and start
playing with the timer subsystem ;). But I guess it's best for now. Let
me have a look...

Jan

Patch

diff --git a/qemu-char.c b/qemu-char.c
index b082bae..1ed6d49 100644
--- a/qemu-char.c
+++ b/qemu-char.c
@@ -465,6 +465,9 @@  static void
mux_chr_update_read_handler(CharDriverState *chr)
     d->focus = d->mux_cnt;
     d->mux_cnt++;
     mux_chr_send_event(d, d->focus, CHR_EVENT_MUX_IN);
+    if (chr->opened) {
+        mux_chr_send_event(d, d->focus, CHR_EVENT_OPENED);
+    }
 }