Mailing List Archive

Hi,

On 22/10/2020 02:43, Elliott Mitchell wrote:
> Linux requires UEFI support to be enabled on ARM64 devices. While many
> ARM64 devices lack ACPI, the writing seems to be on the wall of UEFI/ACPI
> potentially taking over. Some common devices may need ACPI table
> support.
>
> Presently I think it is worth removing the dependancy on CONFIG_EXPERT.

The idea behind EXPERT is to gate any feature that is not considered to
be stable/complete enough to be used in production.

I don't consider the ACPI complete because the parsing of the IORT (used
to discover SMMU and GICv3 ITS) is not there yet.

I vaguely remember some issues on system using SMMU (e.g. Thunder-X)
because Dom0 will try to use the IOMMU and this would break PV drivers.

Therefore I think we at least want to consider to hide SMMUs from dom0
before removing EXPERT. Ideally, I would also like the feature to be
tested in Osstest.

The good news is Xen Project already has systems (e.g. Thunder-X,
Softiron) that can supported ACPI. So it should hopefully be a matter to
tell them to boot with ACPI rather than DT.

Cheers,

--
Julien Grall

On Thu, 22 Oct 2020, Julien Grall wrote:
> On 22/10/2020 02:43, Elliott Mitchell wrote:
> > Linux requires UEFI support to be enabled on ARM64 devices. While many
> > ARM64 devices lack ACPI, the writing seems to be on the wall of UEFI/ACPI
> > potentially taking over. Some common devices may need ACPI table
> > support.
> >
> > Presently I think it is worth removing the dependancy on CONFIG_EXPERT.
>
> The idea behind EXPERT is to gate any feature that is not considered to be
> stable/complete enough to be used in production.

Yes, and from that point of view I don't think we want to remove EXPERT
from ACPI yet. However, the idea of hiding things behind EXPERT works
very well for new esoteric features, something like memory introspection
or memory overcommit. It does not work well for things that are actually
required to boot on the platform.

Typically ACPI systems don't come with device tree at all (RPi4 being an
exception), so users don't really have much of a choice in the matter.

From that point of view, it would be better to remove EXPERT from ACPI,
maybe even build ACPI by default, *but* to add a warning at boot saying
something like:

"ACPI support is experimental. Boot using Device Tree if you can."


That would better convey the risks of using ACPI, while at the same time
making it a bit easier for users to boot on their ACPI-only platforms.


> I don't consider the ACPI complete because the parsing of the IORT (used to
> discover SMMU and GICv3 ITS) is not there yet.
>
> I vaguely remember some issues on system using SMMU (e.g. Thunder-X) because
> Dom0 will try to use the IOMMU and this would break PV drivers.

I am not sure why Dom0 using the IOMMU would break PV drivers? Is it
because the pagetable is not properly updated when mapping foreign
pages?


> Therefore I think we at least want to consider to hide SMMUs from dom0 before
> removing EXPERT. Ideally, I would also like the feature to be tested in
> Osstest.
>
> The good news is Xen Project already has systems (e.g. Thunder-X, Softiron)
> that can supported ACPI. So it should hopefully be a matter to tell them to
> boot with ACPI rather than DT.

I agree that we want to keep ACPI "expert/experimental" given its
current state but maybe we can find a better way to carry that message
than to set EXPERT in Kconfig.

And yes, if we wanted to make ACPI less "expert/experimental" we
definitely need some testing in OSSTest and any critical bugs (e.g. PV
drivers not working) addressed.

On Thu, Oct 22, 2020 at 02:17:13PM -0700, Stefano Stabellini wrote:
> On Thu, 22 Oct 2020, Julien Grall wrote:
> > On 22/10/2020 02:43, Elliott Mitchell wrote:
> > > Linux requires UEFI support to be enabled on ARM64 devices. While many
> > > ARM64 devices lack ACPI, the writing seems to be on the wall of UEFI/ACPI
> > > potentially taking over. Some common devices may need ACPI table
> > > support.
> > >
> > > Presently I think it is worth removing the dependancy on CONFIG_EXPERT.
> >
> > The idea behind EXPERT is to gate any feature that is not considered to be
> > stable/complete enough to be used in production.
>
> Yes, and from that point of view I don't think we want to remove EXPERT
> from ACPI yet. However, the idea of hiding things behind EXPERT works
> very well for new esoteric features, something like memory introspection
> or memory overcommit. It does not work well for things that are actually
> required to boot on the platform.
>
> Typically ACPI systems don't come with device tree at all (RPi4 being an
> exception), so users don't really have much of a choice in the matter.
>
> >From that point of view, it would be better to remove EXPERT from ACPI,
> maybe even build ACPI by default, *but* to add a warning at boot saying
> something like:
>
> "ACPI support is experimental. Boot using Device Tree if you can."
>
>
> That would better convey the risks of using ACPI, while at the same time
> making it a bit easier for users to boot on their ACPI-only platforms.

This matches my view. I was thinking about including "default y", but I
felt the chances of that getting through were lower. I concur with a
warning on boot being a good approach.


--
(\___(\___(\______ --=> 8-) EHM <=-- ______/)___/)___/)
\BS ( | ehem+sigmsg@m5p.com PGP 87145445 | ) /
\_CS\ | _____ -O #include <stddisclaimer.h> O- _____ | / _/
8A19\___\_|_/58D2 7E3D DDF4 7BA6 <-PGP-> 41D1 B375 37D0 8714\_|_/___/5445

Hi Stefano,

On 22/10/2020 22:17, Stefano Stabellini wrote:
> On Thu, 22 Oct 2020, Julien Grall wrote:
>> On 22/10/2020 02:43, Elliott Mitchell wrote:
>>> Linux requires UEFI support to be enabled on ARM64 devices. While many
>>> ARM64 devices lack ACPI, the writing seems to be on the wall of UEFI/ACPI
>>> potentially taking over. Some common devices may need ACPI table
>>> support.
>>>
>>> Presently I think it is worth removing the dependancy on CONFIG_EXPERT.
>>
>> The idea behind EXPERT is to gate any feature that is not considered to be
>> stable/complete enough to be used in production.
>
> Yes, and from that point of view I don't think we want to remove EXPERT
> from ACPI yet. However, the idea of hiding things behind EXPERT works
> very well for new esoteric features, something like memory introspection
> or memory overcommit.

Memaccess is not very new ;).

> It does not work well for things that are actually
> required to boot on the platform.

I am not sure where is the problem. It is easy to select EXPERT from the
menuconfig. It also hints the user that the feature may not fully work.

>
> Typically ACPI systems don't come with device tree at all (RPi4 being an
> exception), so users don't really have much of a choice in the matter.

And they typically have IOMMUs.

>
> From that point of view, it would be better to remove EXPERT from ACPI,
> maybe even build ACPI by default, *but* to add a warning at boot saying
> something like:
>
> "ACPI support is experimental. Boot using Device Tree if you can."
>
>
> That would better convey the risks of using ACPI, while at the same time
> making it a bit easier for users to boot on their ACPI-only platforms.

Right, I agree that this make easier for users to boot Xen on ACPI-only
platform. However, based on above, it is easy enough for a developper to
rebuild Xen with ACPI and EXPERT enabled.

So what sort of users are you targeting?

I am sort of okay to remove EXPERT. But I still think building ACPI by
default is still wrong because our default .config is meant to be
(security) supported. I don't think ACPI can earn this qualification today.

In order to remove EXPERT, there are a few things to needs to be done
(or checked):
1) SUPPORT.MD has a statement about ACPI on Arm
2) DT is favored over ACPI if the two firmware tables are present.

>
>
>> I don't consider the ACPI complete because the parsing of the IORT (used to
>> discover SMMU and GICv3 ITS) is not there yet.
>>
>> I vaguely remember some issues on system using SMMU (e.g. Thunder-X) because
>> Dom0 will try to use the IOMMU and this would break PV drivers.
>
> I am not sure why Dom0 using the IOMMU would break PV drivers? Is it
> because the pagetable is not properly updated when mapping foreign
> pages?

IIRC, yes. This would need to be tested again.

Cheers,

--
Julien Grall

On Fri, 23 Oct 2020, Julien Grall wrote:
> Hi Stefano,
>
> On 22/10/2020 22:17, Stefano Stabellini wrote:
> > On Thu, 22 Oct 2020, Julien Grall wrote:
> > > On 22/10/2020 02:43, Elliott Mitchell wrote:
> > > > Linux requires UEFI support to be enabled on ARM64 devices. While many
> > > > ARM64 devices lack ACPI, the writing seems to be on the wall of
> > > > UEFI/ACPI
> > > > potentially taking over. Some common devices may need ACPI table
> > > > support.
> > > >
> > > > Presently I think it is worth removing the dependancy on CONFIG_EXPERT.
> > >
> > > The idea behind EXPERT is to gate any feature that is not considered to be
> > > stable/complete enough to be used in production.
> >
> > Yes, and from that point of view I don't think we want to remove EXPERT
> > from ACPI yet. However, the idea of hiding things behind EXPERT works
> > very well for new esoteric features, something like memory introspection
> > or memory overcommit.
>
> Memaccess is not very new ;).
>
> > It does not work well for things that are actually
> > required to boot on the platform.
>
> I am not sure where is the problem. It is easy to select EXPERT from the
> menuconfig. It also hints the user that the feature may not fully work.
>
> >
> > Typically ACPI systems don't come with device tree at all (RPi4 being an
> > exception), so users don't really have much of a choice in the matter.
>
> And they typically have IOMMUs.
>
> >
> > From that point of view, it would be better to remove EXPERT from ACPI,
> > maybe even build ACPI by default, *but* to add a warning at boot saying
> > something like:
> >
> > "ACPI support is experimental. Boot using Device Tree if you can."
> >
> >
> > That would better convey the risks of using ACPI, while at the same time
> > making it a bit easier for users to boot on their ACPI-only platforms.
>
> Right, I agree that this make easier for users to boot Xen on ACPI-only
> platform. However, based on above, it is easy enough for a developper to
> rebuild Xen with ACPI and EXPERT enabled.
>
> So what sort of users are you targeting?

Somebody trying Xen for the first time, they might know how to build it
but they might not know that ACPI is not available by default, and they
might not know that they need to enable EXPERT in order to get the ACPI
option in the menu. It is easy to do once you know it is there,
otherwise one might not know where to look in the menu.


> I am sort of okay to remove EXPERT.

OK. This would help (even without building it by default) because as you
go and look at the menu the first time, you'll find ACPI among the
options right away.


> But I still think building ACPI by default
> is still wrong because our default .config is meant to be (security)
> supported. I don't think ACPI can earn this qualification today.

Certainly we don't want to imply ACPI is security supported. I was
looking at SUPPORT.md and it is only says:

"""
EXPERT and DEBUG Kconfig options are not security supported. Other
Kconfig options are supported, if the related features are marked as
supported in this document.
"""

So technically we could enable ACPI in the build by default as ACPI for
ARM is marked as experimental. However, I can see that it is not a
great idea to enable by default an unsupported option in the kconfig, so
from that point of view it might be best to leave ACPI disabled by
default. Probably the best compromise at this time.



> In order to remove EXPERT, there are a few things to needs to be done (or
> checked):
> 1) SUPPORT.MD has a statement about ACPI on Arm

### Host ACPI (via Domain 0)

Status, x86 PV: Supported
Status, ARM: Experimental



> 2) DT is favored over ACPI if the two firmware tables are present.

Good idea. xen/arch/arm/acpi/boot.c:acpi_boot_table_init has:

/*
* Enable ACPI instead of device tree unless
* - ACPI has been disabled explicitly (acpi=off), or
* - the device tree is not empty (it has more than just a /chosen node)
* and ACPI has not been force enabled (acpi=force)
*/
if ( param_acpi_off)
goto disable;
if ( !param_acpi_force &&
device_tree_for_each_node(device_tree_flattened, 0,
dt_scan_depth1_nodes, NULL) )
goto disable;

We should be fine.

Hi Stefano,

On 23/10/2020 17:57, Stefano Stabellini wrote:
> On Fri, 23 Oct 2020, Julien Grall wrote:
>> Hi Stefano,
>>
>> On 22/10/2020 22:17, Stefano Stabellini wrote:
>>> On Thu, 22 Oct 2020, Julien Grall wrote:
>>>> On 22/10/2020 02:43, Elliott Mitchell wrote:
>>>>> Linux requires UEFI support to be enabled on ARM64 devices. While many
>>>>> ARM64 devices lack ACPI, the writing seems to be on the wall of
>>>>> UEFI/ACPI
>>>>> potentially taking over. Some common devices may need ACPI table
>>>>> support.
>>>>>
>>>>> Presently I think it is worth removing the dependancy on CONFIG_EXPERT.
>>>>
>>>> The idea behind EXPERT is to gate any feature that is not considered to be
>>>> stable/complete enough to be used in production.
>>>
>>> Yes, and from that point of view I don't think we want to remove EXPERT
>>> from ACPI yet. However, the idea of hiding things behind EXPERT works
>>> very well for new esoteric features, something like memory introspection
>>> or memory overcommit.
>>
>> Memaccess is not very new ;).
>>
>>> It does not work well for things that are actually
>>> required to boot on the platform.
>>
>> I am not sure where is the problem. It is easy to select EXPERT from the
>> menuconfig. It also hints the user that the feature may not fully work.
>>
>>>
>>> Typically ACPI systems don't come with device tree at all (RPi4 being an
>>> exception), so users don't really have much of a choice in the matter.
>>
>> And they typically have IOMMUs.
>>
>>>
>>> From that point of view, it would be better to remove EXPERT from ACPI,
>>> maybe even build ACPI by default, *but* to add a warning at boot saying
>>> something like:
>>>
>>> "ACPI support is experimental. Boot using Device Tree if you can."
>>>
>>>
>>> That would better convey the risks of using ACPI, while at the same time
>>> making it a bit easier for users to boot on their ACPI-only platforms.
>>
>> Right, I agree that this make easier for users to boot Xen on ACPI-only
>> platform. However, based on above, it is easy enough for a developper to
>> rebuild Xen with ACPI and EXPERT enabled.
>>
>> So what sort of users are you targeting?
>
> Somebody trying Xen for the first time, they might know how to build it
> but they might not know that ACPI is not available by default, and they
> might not know that they need to enable EXPERT in order to get the ACPI
> option in the menu. It is easy to do once you know it is there,
> otherwise one might not know where to look in the menu.

Right, EXPERT can now be enabled using Kconfig. So it is not very
different from an option Foo has been hidden because the dependency Bar
has not been selected.

It should be easy enough (if it is not we should fix it) to figure out
the dependency when searching the option via menuconfig.

>
>
>> I am sort of okay to remove EXPERT.
>
> OK. This would help (even without building it by default) because as you
> go and look at the menu the first time, you'll find ACPI among the
> options right away.

To be honest, this step is probably the easiest in the full process to
get Xen build and booted on Arm.

I briefly looked at Elliot's v2, and I can't keep thinking that we are
trying to re-invent EXPERT for ACPI because we think the feature is
*more* important than any other feature gated by EXPERT.

In fact, all the features behind EXPERT are important. But they have
been gated by EXPERT because they are not mature enough.

We already moved EXPERT from a command line option to a menuconfig
option. So it should be easy enough to enable it now. If it still not
the case, then we should improve it.

But I don't think ACPI is mature enough to deserve a different
treatment. It would be more useful to get to the stage where ACPI can
work without any crash/issue first.

>
>
>> But I still think building ACPI by default
>> is still wrong because our default .config is meant to be (security)
>> supported. I don't think ACPI can earn this qualification today.
>
> Certainly we don't want to imply ACPI is security supported. I was
> looking at SUPPORT.md and it is only says:
>
> """
> EXPERT and DEBUG Kconfig options are not security supported. Other
> Kconfig options are supported, if the related features are marked as
> supported in this document.
> """
>
> So technically we could enable ACPI in the build by default as ACPI for
> ARM is marked as experimental. However, I can see that it is not a
> great idea to enable by default an unsupported option in the kconfig, so
> from that point of view it might be best to leave ACPI disabled by
> default. Probably the best compromise at this time.
From my understanding, the goal of EXPERT was to gate such features.
With your suggestion, it is not clear to me what's the difference
between "experimental" and option gated by "EXPERT".

Do you mind clarifying?

Cheers,

--
Julien Grall

On Mon, 26 Oct 2020, Julien Grall wrote:
> Hi Stefano,
>
> On 23/10/2020 17:57, Stefano Stabellini wrote:
> > On Fri, 23 Oct 2020, Julien Grall wrote:
> > > Hi Stefano,
> > >
> > > On 22/10/2020 22:17, Stefano Stabellini wrote:
> > > > On Thu, 22 Oct 2020, Julien Grall wrote:
> > > > > On 22/10/2020 02:43, Elliott Mitchell wrote:
> > > > > > Linux requires UEFI support to be enabled on ARM64 devices. While
> > > > > > many
> > > > > > ARM64 devices lack ACPI, the writing seems to be on the wall of
> > > > > > UEFI/ACPI
> > > > > > potentially taking over. Some common devices may need ACPI table
> > > > > > support.
> > > > > >
> > > > > > Presently I think it is worth removing the dependancy on
> > > > > > CONFIG_EXPERT.
> > > > >
> > > > > The idea behind EXPERT is to gate any feature that is not considered
> > > > > to be
> > > > > stable/complete enough to be used in production.
> > > >
> > > > Yes, and from that point of view I don't think we want to remove EXPERT
> > > > from ACPI yet. However, the idea of hiding things behind EXPERT works
> > > > very well for new esoteric features, something like memory introspection
> > > > or memory overcommit.
> > >
> > > Memaccess is not very new ;).
> > >
> > > > It does not work well for things that are actually
> > > > required to boot on the platform.
> > >
> > > I am not sure where is the problem. It is easy to select EXPERT from the
> > > menuconfig. It also hints the user that the feature may not fully work.
> > >
> > > >
> > > > Typically ACPI systems don't come with device tree at all (RPi4 being an
> > > > exception), so users don't really have much of a choice in the matter.
> > >
> > > And they typically have IOMMUs.
> > >
> > > >
> > > > From that point of view, it would be better to remove EXPERT from
> > > > ACPI,
> > > > maybe even build ACPI by default, *but* to add a warning at boot saying
> > > > something like:
> > > >
> > > > "ACPI support is experimental. Boot using Device Tree if you can."
> > > >
> > > >
> > > > That would better convey the risks of using ACPI, while at the same time
> > > > making it a bit easier for users to boot on their ACPI-only platforms.
> > >
> > > Right, I agree that this make easier for users to boot Xen on ACPI-only
> > > platform. However, based on above, it is easy enough for a developper to
> > > rebuild Xen with ACPI and EXPERT enabled.
> > >
> > > So what sort of users are you targeting?
> >
> > Somebody trying Xen for the first time, they might know how to build it
> > but they might not know that ACPI is not available by default, and they
> > might not know that they need to enable EXPERT in order to get the ACPI
> > option in the menu. It is easy to do once you know it is there,
> > otherwise one might not know where to look in the menu.
>
> Right, EXPERT can now be enabled using Kconfig. So it is not very different
> from an option Foo has been hidden because the dependency Bar has not been
> selected.
>
> It should be easy enough (if it is not we should fix it) to figure out the
> dependency when searching the option via menuconfig.

I do `make menuconfig` and there is no ACPI option. How do I even know
that ACPI has a kconfig option to enable? I'd assume that ACPI is always
enabled in the kconfig unless told otherwise.

But let's say that you know that you need to look for ACPI. I'd use the
search function, and it tells me:

Symbol: ACPI [=n] ?
Type : bool ?
Prompt: ACPI (Advanced Configuration and Power Interface) Support ?
Location: ?
(1) -> Architecture Features ?
Defined at arch/arm/Kconfig:34 ?
Depends on: ARM_64 [=y]

I go and look "Architecture Features" as told, but it is not there. How
do I need that I need to enable "Configure standard Xen features (expert
users)" to get that option?


> > > I am sort of okay to remove EXPERT.
> >
> > OK. This would help (even without building it by default) because as you
> > go and look at the menu the first time, you'll find ACPI among the
> > options right away.
>
> To be honest, this step is probably the easiest in the full process to get Xen
> build and booted on Arm.
>
> I briefly looked at Elliot's v2, and I can't keep thinking that we are trying
> to re-invent EXPERT for ACPI because we think the feature is *more* important
> than any other feature gated by EXPERT.
>
> In fact, all the features behind EXPERT are important. But they have been
> gated by EXPERT because they are not mature enough.

It is not as much a matter of "importance" but a matter of required for
booting. I don't think that EXPERT is really a good tool for gating
things that are required for booting. If we had something else (not
ACPI) that is required for booting and marked as EXPERT, I'd say the
same thing. The only other thing that might qualify is ITS support.


> We already moved EXPERT from a command line option to a menuconfig option. So
> it should be easy enough to enable it now. If it still not the case, then we
> should improve it.
>
> But I don't think ACPI is mature enough to deserve a different treatment.

I fully agree ACPI is not mature.


> It would be more useful to get to the stage where ACPI can work
> without any crash/issue first.

Yes indeed. I don't have any stake in this, given that none of my
systems have ACPI, so I'd better shut up maybe :-)


> > > But I still think building ACPI by default
> > > is still wrong because our default .config is meant to be (security)
> > > supported. I don't think ACPI can earn this qualification today.
> >
> > Certainly we don't want to imply ACPI is security supported. I was
> > looking at SUPPORT.md and it is only says:
> >
> > """
> > EXPERT and DEBUG Kconfig options are not security supported. Other
> > Kconfig options are supported, if the related features are marked as
> > supported in this document.
> > """
> >
> > So technically we could enable ACPI in the build by default as ACPI for
> > ARM is marked as experimental. However, I can see that it is not a
> > great idea to enable by default an unsupported option in the kconfig, so
> > from that point of view it might be best to leave ACPI disabled by
> > default. Probably the best compromise at this time.
> From my understanding, the goal of EXPERT was to gate such features. With your
> suggestion, it is not clear to me what's the difference between "experimental"
> and option gated by "EXPERT".
>
> Do you mind clarifying?

Ah! That's a good question actually. Is the expectation and
"experimental" in SUPPORT.md and EXPERT in Kconfig are pretty much the
same thing? I didn't think so. Let's have a look. SUPPORT.md says:

### KCONFIG Expert

Status: Experimental


And EXPERT says:

config EXPERT
bool "Configure standard Xen features (expert users)"
help
This option allows certain base Xen options and settings
to be disabled or tweaked. This is for specialized environments
which can tolerate a "non-standard" Xen.
Only use this if you really know what you are doing.
Xen binaries built with this option enabled are not security
supported.


It seems that they are not the same: EXPERT is a *subset* of
Experimental for certain Xen options "for specialized environments" and
"expert users, which I think makes sense. ACPI is a good example of
something "experimental" but not for specialized environments.

On Mon, Oct 26, 2020 at 09:19:49AM +0000, Julien Grall wrote:
> On 23/10/2020 17:57, Stefano Stabellini wrote:
> > On Fri, 23 Oct 2020, Julien Grall wrote:

> >> I am sort of okay to remove EXPERT.
> >
> > OK. This would help (even without building it by default) because as you
> > go and look at the menu the first time, you'll find ACPI among the
> > options right away.
>
> To be honest, this step is probably the easiest in the full process to
> get Xen build and booted on Arm.
>
> I briefly looked at Elliot's v2, and I can't keep thinking that we are
> trying to re-invent EXPERT for ACPI because we think the feature is
> *more* important than any other feature gated by EXPERT.

Yet might that statement in fact be true?

Most of the features currently controlled by CONFIG_EXPERT are relatively
small tweaks which enable less often used features. Some of those are
very high value in certain environments, but they're unimportant in
common environments. Changing the scheduler might get you an extra
10-50% performance improvement in a special environment.

ACPI support isn't like this. I'm unaware what Masami Hiramatsu's system
does if a CONFIG_ACPI=n Xen build is tried. I haven't actually tried
such a build on mine, but from the code it looks like Xen would panic if
built that way. No output of any sort. Simply panic with the device
appearing to go inert.

> In fact, all the features behind EXPERT are important. But they have
> been gated by EXPERT because they are not mature enough.

> But I don't think ACPI is mature enough to deserve a different
> treatment. It would be more useful to get to the stage where ACPI can
> work without any crash/issue first.

The difference is the severity of failure if the option is off, but needs
to be on. Most CONFIG_EXPERT options will merely be performance
reductions or security situations unacceptable to some users.
CONFIG_ACPI=n on the wrong system could be a panic with *no* output.


Tainting sounds reasonable. Messages in `dmesg` make sense. A message
plus 10 second pause on boot seems reasonable. I think if CONFIG_ACPI is
off, there should be code to try to detect ACPI and emit a warning if
anything is detected.


--
(\___(\___(\______ --=> 8-) EHM <=-- ______/)___/)___/)
\BS ( | ehem+sigmsg@m5p.com PGP 87145445 | ) /
\_CS\ | _____ -O #include <stddisclaimer.h> O- _____ | / _/
8A19\___\_|_/58D2 7E3D DDF4 7BA6 <-PGP-> 41D1 B375 37D0 8714\_|_/___/5445

Hi Elliott,

On 04/11/2020 19:03, Elliott Mitchell wrote:
> On Mon, Oct 26, 2020 at 09:19:49AM +0000, Julien Grall wrote:
>> On 23/10/2020 17:57, Stefano Stabellini wrote:
>>> On Fri, 23 Oct 2020, Julien Grall wrote:
>
>>>> I am sort of okay to remove EXPERT.
>>>
>>> OK. This would help (even without building it by default) because as you
>>> go and look at the menu the first time, you'll find ACPI among the
>>> options right away.
>>
>> To be honest, this step is probably the easiest in the full process to
>> get Xen build and booted on Arm.
>>
>> I briefly looked at Elliot's v2, and I can't keep thinking that we are
>> trying to re-invent EXPERT for ACPI because we think the feature is
>> *more* important than any other feature gated by EXPERT.
>
> Yet might that statement in fact be true?

Everyone has a different opinion on what's important or not. I am sure
we can spend hours bikeshedding on that...

>
> Most of the features currently controlled by CONFIG_EXPERT are relatively
> small tweaks which enable less often used features. Some of those are
> very high value in certain environments, but they're unimportant in
> common environments. Changing the scheduler might get you an extra
> 10-50% performance improvement in a special environment.
>
> ACPI support isn't like this. I'm unaware what Masami Hiramatsu's system
> does if a CONFIG_ACPI=n Xen build is tried. I haven't actually tried
> such a build on mine, but from the code it looks like Xen would panic if
> built that way. No output of any sort. Simply panic with the device
> appearing to go inert.
There will always be cases where the console is not working:
1) There is no driver in Xen
2) There is no SPCR table present

I think you are in the second situation and you had to enable
earlyprintk. Is that correct?

>> In fact, all the features behind EXPERT are important. But they have
>> been gated by EXPERT because they are not mature enough.
>
>> But I don't think ACPI is mature enough to deserve a different
>> treatment. It would be more useful to get to the stage where ACPI can
>> work without any crash/issue first.
>
> The difference is the severity of failure if the option is off, but needs
> to be on. Most CONFIG_EXPERT options will merely be performance
> reductions or security situations unacceptable to some users.
> CONFIG_ACPI=n on the wrong system could be a panic with *no* output. >
>
> Tainting sounds reasonable. Messages in `dmesg` make sense. A message
> plus 10 second pause on boot seems reasonable. I think if CONFIG_ACPI is
> off, there should be code to try to detect ACPI and emit a warning if
> anything is detected.

All of this would be moot if we focus on getting ACPI (or just a subset)
working on a few platforms (e.g. RPI, Thunder-X).

I don't think we are too far from this acheviement. This would be IMHO
be enough to move ACPI one way up in the support "ladder". We might even
be able to do this for 4.15.

Cheers,

--
Julien Grall