Re: [PATCH] xen-pciback: allow compiling on other archs than x86

[Oleksandr Andrushchenko]

On 21.09.21 23:44, Stefano Stabellini wrote:
> On Tue, 21 Sep 2021, Oleksandr Andrushchenko wrote:
>> On 21.09.21 10:09, Juergen Gross wrote:
>>> On 21.09.21 09:00, Oleksandr Andrushchenko wrote:
>>>> On 21.09.21 09:49, Juergen Gross wrote:
>>>>> On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
>>>>>> On 21.09.21 09:07, Juergen Gross wrote:
>>>>>>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>>>>>>> On 21.09.21 08:20, Juergen Gross wrote:
>>>>>>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>>>>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>>>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>>>>>>>> Hello, Stefano!
>>>>>>>>>>>>>
>>>>>>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>>>>>>>> Hi Oleksandr,
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>>>>>>>> Not only that
>>>>>>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>>>>>>>> the same time.
>>>>>>>>>>>>> Correct, it is not used
>>>>>>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>>>>>>>
>>>>>>>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>>>>>>>
>>>>>>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>>>>>>>>           pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>>>>>>>>           toolstack needs to know which PCI devices can be passed through it reads
>>>>>>>>>>>>>           that from the relevant sysfs entries of the pciback.
>>>>>>>>>>>>>
>>>>>>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>>>>>>>>           a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>>>>>>>>           to pciback (strictly speaking it is not required that the device is bound to
>>>>>>>>>>>>>           pciback, but pciback is again used as a database of the passed through PCI
>>>>>>>>>>>>>           devices, so we can re-bind the devices back to their original drivers when
>>>>>>>>>>>>>           guest domain shuts down)
>>>>>>>>>>>>>
>>>>>>>>>>>>> 3. Device reset
>>>>>>>>>>>>>
>>>>>>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Please see [1] and [2]:
>>>>>>>>>>>>>
>>>>>>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>>>>>>>
>>>>>>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>>>>>>>> the rest like vPCI etc.
>>>>>>>>>>>>>
>>>>>>>>>>>>> 3. pcifront is not used on Arm
>>>>>>>>>>>> It is neither in x86 PVH/HVM guests.
>>>>>>>>>>> Didn't know that, thank you for pointing
>>>>>>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>>>>>>>> all the goals above.
>>>>>>>>>>>>>
>>>>>>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>>>>>>>> be supported.
>>>>>>>>>>> Agree, that the final solution should have the driver split
>>>>>>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>>>>>>>> is done first.
>>>>>>>>>>>>
>>>>>>>>>>>> I don't mind doing it in either sequence.
>>>>>>>>>>>>
>>>>>>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>>>>>>>
>>>>>>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>>>>>>>
>>>>>>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>>>>>>>
>>>>>>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>>>>>>>
>>>>>>>>>>> have enough bandwidth for that piece of work at the moment.
>>>>>>>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>>>>>>>> an enormous rework. At the same time I don't think we should enable the
>>>>>>>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>>>>>>> As the first stage before the driver is split or ifdef's used - can we take the patch
>>>>>>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>>>>>>>> for other architectures and common code move.
>>>>>>> Fine with me in principle. I need to take a more thorough look
>>>>>>> at the patch, though.
>>>>>> Of course
>>>>>>>>>> I am wonder if there is a simple:
>>>>>>>>>>
>>>>>>>>>> if (!xen_pv_domain())
>>>>>>>>>>          return;
>>>>>>>>>>
>>>>>>>>>> That we could add in a couple of places in pciback to stop it from
>>>>>>>>>> initializing the parts we don't care about. Something along these lines
>>>>>>>>>> (untested and probably incomplete).
>>>>>>>>>>
>>>>>>>>>> What do you guys think?
>>>>>>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>>>>>>>> with dom0 running as PVH. I don't think this is working right now, but
>>>>>>>>> adding more code making it even harder to work should be avoided.
>>>>>>>>>
>>>>>>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>>>>>>>> index da34ce85dc88..991ba0a9b359 100644
>>>>>>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>>>>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>>>>>>>> @@ -15,6 +15,7 @@
>>>>>>>>>>       #include <xen/xenbus.h>
>>>>>>>>>>       #include <xen/events.h>
>>>>>>>>>>       #include <xen/pci.h>
>>>>>>>>>> +#include <xen/xen.h>
>>>>>>>>>>       #include "pciback.h"
>>>>>>>>>>         #define INVALID_EVTCHN_IRQ  (-1)
>>>>>>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>>>>>>>>                       const struct xenbus_device_id *id)
>>>>>>>>>>       {
>>>>>>>>>>           int err = 0;
>>>>>>>>>> -    struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>>>>>>>> +    struct xen_pcibk_device *pdev;
>>>>>>>>>> +
>>>>>>>>>> +    if (!xen_pv_domain())
>>>>>>>>>> +        return 0;
>>>>>>>>>>       +    pdev = alloc_pdev(dev);
>>>>>>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>>>>>>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>>>>>>>
>>>>>>>>>>           if (pdev == NULL) {
>>>>>>>>>>               err = -ENOMEM;
>>>>>>>>>>               xenbus_dev_fatal(dev, err,
>>>>>>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>>>>>>>>         int __init xen_pcibk_xenbus_register(void)
>>>>>>>>>>       {
>>>>>>>>>> +    if (!xen_pv_domain())
>>>>>>>>>> +        return 0;
>>>>>>>>>> +
>>>>>>>>> Use #ifdef CONFIG_X86 instead.
>>>>>>>> The title of this patch says that we want to allow this driver for other archs
>>>>>>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>>>>>>>> right with that respect. Instead, we may want having something like a
>>>>>>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>>>>>>>> or something which is architecture agnostic.
>>>>>>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
>>>>>>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
>>>>>>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
>>>>>>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
>>>>>>> wanted. The splitting of the driver can still be done later.
>>>>>> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
>>>>>> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
>>>>>> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
>>>>>> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
>>>>>> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
>>>>> No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
>>>>> is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
>>>>> not set (this will be the case on Arm).
>>>> But this will require that the existing kernel configurations out there have to additionally define CONFIG_XEN_PCI_STUB to get what they had before with simply enabling CONFIG_XEN_PCIDEV_BACKEND. My point was that it is probably desirable not to break
>>>> the things while doing the split/re-work.
>>> By letting XEN_PCIDEV_BACKEND select XEN_PCI_STUB this won't happen.
>> Indeed
>>>> I also thought that "compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is not set"
>>>> may have more code gated rather than with gating unwanted code with CONFIG_XEN_PCI_STUB.
>>>> I am not quite sure about this though.
>>> This would be a very weird semantics of XEN_PCI_STUB, as the stub part
>>> is needed on X86 and on Arm.
>>>
>>> Gating could even be done with Stefano's patch just by replacing his
>>> "!xen_pv_domain()" tests with "!IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND)".
>> Makes sense.
>>
>> Another question if we do not want the code to be compiled or not executed?
>>
>> If the later then we can define something like:
>>
>> bool need_pv_part(void)
>>
>> {
>>
>>       return IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND);
>>
>> }
>>
>> and then just use need_pv_part() for the checks where it is needed.
>>
>> This allows avoiding multiple ifdef's through the code
> This is even better.
>
> For my clarity, Oleksandr, are you OK with adding a few need_pv_part()
> checks through the code as part of this series so that the PV PCI
> backend is not initialized?
Yes
>
> I don't have a good test environment ready for this, so I cannot really
> volunteer myself.
>
> I would prefer if we made this change as part of this series so that the
> PV PCI backend features doesn't get enabled on ARM, not even temporarily.
Ok, I will push v2 today with the additional patch for PV