On 4/20/20 1:55 PM, Pete Batard wrote:
On 2020.04.20 07:45, Ard Biesheuvel wrote:
...
I rejected ACPI 6.3 table upgrades because, in their current form, the
only thing they achieve is losing the ability to boot an OS that
predates ACPI 6.3. Every piece of the platform currently being
described via ACPI can be described using 5.1 tables.
Actually, that is not technically correct.
But I will concede that you probably aren't aware of this, because we
didn't bring up this point when we discussed the matter earlier.
One thing you may want to know is that most of the binary blobs we
picked up from the Microsoft IoT endeavour were ACPI 6.0, and,
especially, the MADT we got had actually been abusing the GICR Base
Address field, which was introduced in 6.0, to place then "nonsensical"
value of 1 in there.
And for reasons that only Microsoft knows, unless you do have a value of
1 there, Windows 10 on the Pi 3 will not boot at all. In other words, if
you are to use ACPI 6.x with carry-over defaults from 5.1, Windows 10
can simply not boot, at least on the Pi 3.
Now, what happened is that, when we started introducing some 5.1 tables
elsewhere, I went with trying to harmonize everything to 5.1 and
effectively *downgraded* the 6.0 tables we got from Microsoft to 5.1
(which nobody seemed to mind then, on account, I will also assert, that
nobody realized that we actually had 6.0 tables where 6.0 features might
be relevant).
It turns out that, if you do use 5.1, Windows 10 *seems* to be happy
about the MADT, even as it is obviously missing a GICR Base Address
field with a 1 value, and *appears* to work as expected.
However, because of the proprietary nature of the OS, we have absolutely
no idea whether having a "properly manipulated" GICR Base Address for
Windows is important or not. The obvious reasoning is that, if Microsoft
abused that field in their tables and especially if, when using ACPI
6.x, Windows 10 doesn't boot unless it sees the expected value there, an
ACPI 6.x MADT with the relevant value populated might be something that
we actually want to see.
I agree that it makes sense to harmonize the table versions, and avoid
mixing and matching different versions of the spec. The structures are
usually defined in a way where new fields added at then end, and the
sizes are explicitly defined, so that an older OS can consume newer
versions of the tables.
At least, that is the theory ...
ACPI 6.3 tables are not 'better' or 'more current' than 5.1 ones.
The example above begs to differ (though of course, short of actually
knowing why the heck Microsoft decided to abuse GICR Base Address, it's
hard to know what 'better' means in this case).
It does seem to me like Microsoft were effectively using the most recent
ACPI version they could use when they produced their IoT tables. And we
certainly did happen to downgrade some of that, and may have gotten
lucky that it all seemed to work (with the point being that, if we seem
to be happy that downgrading ACPI tables from 6.0 to 5.1 and validating
that we aren't observing obvious ill effects is okay, then upgrading
ACPI tables to 6.3 and validating that we aren't observing ill effects
should be fine too).
If compatibility with Microsoft Windows 10 requires a certain minimum
version of the tables, then I don't mind adhering to that.
ACPI 6.3 tables should be used if/when they are needed, and not before.
I disagree here again, on account of trying to showcase elements where
we can, and this is one place we can do this.
I see absolutely zero technical reasons to want to stick with ACPI 5.1
especially as we already have one 6.3 table (PPTT) and we're going to
have to upgrade a couple others to at least 6.0 anyway to get back to
the position we were in with the Microsoft's blobs, so that we can be
more confident about Windows' behaviour.
As a matter of fact, using ACPI 6.3 before it is effectively needed is
exactly the kind of move I see as wanting to apply when we know that a
platform is going to be used as a de facto showcase. It may seem like a
simple "newer is always better!" push to you, but the way I see it is
that, by doing so, we are going to help developers of new platforms, who
will be looking at the Pi for reference (whether we like it or not) and,
if they are developing for modern hardware, are going to be looking at
using new features that introduced only in recent ACPI. Thus, even if we
don't make any use of these features on the Pi's, those developers might
be grateful to find that there exists a working example of using a
relatively up to date ACPI, and that they won't have to reinvent the wheel.
Especially, as opposed to what might be the case for other platforms, we
don't have a baggage of "older" OSes that we may hinder support for if
we do move to ACPI 6.3 (especially as, outside of Windows, most Pi 3
OSes will be using DT rather than ACPI, and, without SD and Genet
support, the idea of older OSes needing to work with Pi 4 is a bit
ludicrous).
So, whereas I agree that your reasoning is generally sound for most
platforms, the fact that the Pi is going to be used as a de-facto
showcase or starting point by folks interested in developing a new ARM
platform, makes, I will assert, the situation a bit different here, and
I would really appreciate if you could give some more thoughts to the
points I am trying to make. In this case, I believe that the Pi should
be the "exception that proves the rule".
Furthermore, and this IMO is the most important point, we are not going
to be sitting idle if we do upgrade to 6.3 and someone comes to us with
a report that using modern ACPI appears to be causing them trouble
(which, I will also assert, may actually be something that might be of
great interest to folks on this list if that happens). As opposed to
proprietary platforms, where ACPI is pretty much set in stone by the
vendor and where addressing issues is a struggle, and, even if the
vendor is very reactive and produces an update, where flashing a new
firmware can be a tricky operation for some users, we are Open Source,
easy to reach for reports, and the firmware "update" process of the Pi
could not be any simpler (copy a file to an SD card and you're done). So
I'd really like to help squelch the idea that if we do decide to upgrade
to ACPI 6.3, and we effectively find out that it causes issues, we're
simply going to stand by and let these unaddressed, whilst telling users
of the platform "too bad". None of what we are doing is ever meant to be
static or immutable, even after a goal has been reached. Thus, if there
are issues with upgrading to 6.3, we're going to make darn sure that we
address them, in the same way we have been planning to address issues
that might be raised from some of our forced downgrades from 6.0 to 5.1.
Thanks Pete.
I still don't agree with the whole showcase aspect of this discussion,
but since there are apparently other good reasons to upgrade these
tables that nobody brought up before, I am not going to oppose any
changes on this front, provided that the commit logs articulate in
sufficient detail why we think this is necessary (so that people looking
at this port for guidance have the background as well)
Out of curiosity, why did we need a 6.3 PPTT for RPi4?
-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#57628): https://edk2.groups.io/g/devel/message/57628
Mute This Topic: https://groups.io/mt/73127191/21656
Group Owner: devel+ow...@edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
