From: "Philippe Mathieu-Daudé" <f4bug@amsat.org>
To: Markus Armbruster <armbru@redhat.com>,
Peter Maydell <peter.maydell@linaro.org>,
Eduardo Habkost <ehabkost@redhat.com>
Cc: Damien Hedde <damien.hedde@greensocs.com>,
QEMU Developers <qemu-devel@nongnu.org>
Subject: Re: Resetting non-qdev children in a 3-phase reset device
Date: Sat, 24 Apr 2021 01:28:56 +0200 [thread overview]
Message-ID: <3dcb8efa-8f6b-0f45-a753-cc58d3bf9855@amsat.org> (raw)
In-Reply-To: <2f4155ca-ee84-75f0-8b35-aac6b18ee4f0@amsat.org>
On 4/24/21 1:06 AM, Philippe Mathieu-Daudé wrote:
> On 4/22/21 4:20 PM, Philippe Mathieu-Daudé wrote:
>> On 4/22/21 3:21 PM, Markus Armbruster wrote:
>>> Peter Maydell <peter.maydell@linaro.org> writes:
>
>>> Most qdevs plug into a qbus, but some don't.
>>>
>>> DeviceClass member @bus_type names the kind of bus the device plugs
>>> into. It's a QOM type name. Example: for a PCI device, it's
>>> TYPE_PCI_BUS, and the device must be plugged into an instance of a
>>> (subtype of) TYPE_PCI_BUS.
>>>
>>> If @bus_type is null, the device does not plug into any qbus.
>>>
>>> The qbus a device is plugged into is also called the parent bus. Not to
>>> be confused with the QOM parent.
>>>
>>>>> But even without parent they end in the /unattached
>>>>> container below /machine, so if the reset were there, the
>>>>> machine could still iterate over the /unattached children.
>>>>
>>>> ...yes, /unattached is what I was thinking about.
>>>>
>>>> My current half-thought-through view is that where we ought
>>>> to try to end up is something like:
>>>>
>>>> * "real" buses should continue to propagate reset
>>>> (A "real" bus is like PCI, SCSI, and other buses where the real
>>>> hardware has a concept of a "bus reset" or where the power to the
>>>> plugged device comes from the bus so that powercycling the
>>>> controller naturally powercycles the devices. Sysbus is not a
>>>> "real" bus; I haven't checked the others to see if we have any
>>>> other non-real buses)
>>>> * reset should follow the QOM tree for objects not on a "real" bus
>>>> (that is, the qdev "reset this device" function should do
>>>> "iterate through my QOM children and reset those which are not
>>>> on a real bus" as well as its current "reset myself" and "reset
>>>> every qbus I have")
>>>> * instead of reset starting with the sysbus and working along the
>>>> qbus hierarchy, we start by resetting the machine. That should
>>>> include resetting all the QOM children of the machine. Any
>>>> device which has a qbus should reset the qbus as part of its
>>>> reset, but only "real" buses reset their children when reset.
>>>
>>> Sounds like an approximation of reset wire modelling :)
>>>
>>> In a real machine, the reset signal travels along "wires" (in quotes,
>>> because it need not be a dedicated wire, although it commonly is)
>>>
>>> We're not modelling these wires explicitly so far. Instead, we make
>>> assumptions such as "reset flows along the qdev tree", which are close
>>> enough except when they aren't.
>>>
>>> What you propose is likely closer to reality than what we have now.
>>
>> Then maybe reality is easier to model =)
>>
>>> Do I make sense?
>>
>> I guess so. Now I wonder if Peter's approach is doable while still
>> having "incompletely QOMified devices".
>>
>> But if we can propagate reset tree via QOM, it is a good excuse
>> to finish QOM'ifying devices and enforce the API to prohibit non-QOM
>> ones.
>>
>> And remove the crutch in device_set_realized().
>>
>>>> That means that, for instance, if you reset an SoC container object
>>>> it will reset all the sub-devices within the SoC and the miscellaneous
>>>> bits of glue logic like OR gates it might also own[*]. It also means that
>>>> CPU objects should no longer need weird special casing, because they
>>>> are part of the QOM hierarchy and get reset that way.
>>>>
>>>> [*] Fun fact: TYPE_OR_IRQ inherits directly from TYPE_DEVICE which
>>>> means that pretty much no instances of it ever get reset.
>>>>
>>>> There is of course a massive unsolved problem with this idea, which
>>>> is the usual "how do we get there from here" one.
>>>>
>>>> (Eventually I think we might be able to collapse TYPE_SYS_BUS_DEVICE
>>>> down into TYPE_DEVICE: there is no particular reason why a TYPE_DEVICE
>>>> can have GPIO inputs and outputs but only a TYPE_SYS_BUS_DEVICE can
>>>> claim to have MMIO regions and IRQs. "Only sysbus devices get reset"
>>>> is a big part of why a lot of devices today are sysbus.)
>
> Looking at qemu_register_reset() uses I found this commit:
>
> commit 0c7322cfd3fd382c0096c2a9f00775818a878e13
> Date: Mon Jun 29 08:21:10 2015 +0200
>
> watchdog/diag288: correctly register for system reset requests
>
> The diag288 watchdog is no sysbus device, therefore it doesn't get
> triggered on resets automatically using dc->reset.
>
> Let's register the reset handler manually, so we get correctly notified
> again when a system reset was requested. Also reset the watchdog on
> subsystem resets that don't trigger a full system reset.
>
> Why is the reset() handler in DeviceClass and not in SysbusDeviceClass
> if "Only sysbus devices get reset"? ...
Ah, probably because the problem is generic to all busses (ISA, ...)
and not just sysbus.
>>> Sysbus may habe been a design mistake. It goes back the qdev design
>>> assumption "every device plugs into exactly one bus, every bus is part
>>> of exactly one device, and the main system bus is the root of this
>>> tree". The assumption ceased to hold long ago, but we still have
>>> sysbus.
>
next prev parent reply other threads:[~2021-04-23 23:29 UTC|newest]
Thread overview: 16+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-04-09 18:13 Resetting non-qdev children in a 3-phase reset device Peter Maydell
2021-04-18 20:16 ` Philippe Mathieu-Daudé
2021-04-19 9:03 ` Peter Maydell
2021-04-22 13:21 ` Markus Armbruster
2021-04-22 14:20 ` Philippe Mathieu-Daudé
2021-04-23 23:06 ` Philippe Mathieu-Daudé
2021-04-23 23:28 ` Philippe Mathieu-Daudé [this message]
2021-04-24 5:28 ` Markus Armbruster
2021-04-24 13:04 ` Philippe Mathieu-Daudé
2021-04-24 13:15 ` Philippe Mathieu-Daudé
2021-04-25 18:33 ` Peter Maydell
2021-04-26 5:19 ` Markus Armbruster
2021-04-26 9:09 ` Peter Maydell
2021-04-26 9:23 ` Philippe Mathieu-Daudé
2021-04-26 9:33 ` Peter Maydell
2021-04-26 11:14 ` Philippe Mathieu-Daudé
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=3dcb8efa-8f6b-0f45-a753-cc58d3bf9855@amsat.org \
--to=f4bug@amsat.org \
--cc=armbru@redhat.com \
--cc=damien.hedde@greensocs.com \
--cc=ehabkost@redhat.com \
--cc=peter.maydell@linaro.org \
--cc=qemu-devel@nongnu.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).