On Nov 20, 2008, at 6:35 PM, David Gibson wrote:
On Thu, Nov 20, 2008 at 01:07:11PM -0600, Milton Miller wrote:
David Gibson wrote at 2008-11-18 00:28:28:
On Mon, Nov 17, 2008 at 01:41:24PM -0600, Timur Tabi wrote:
On Thu, Oct 23, 2008 at 9:54 PM, David Gibson
<[EMAIL PROTECTED]> wrote:
This patch adds a new backend for the hvc console based on the
low-level udbg callbacks. This effectively implements a working
runtime console in terms of the simple udbg primitives. This is
kind
of a hack - since udbg isn't something you really want to be using
routinely - but it's really useful during bringup.
Why is udbg hvc something you want to not use routinely?
Well.. because the udbg methods themselves are not sometimes (not
always, but reasonably often), something you don't really want to use
except during bringup. A number of them use various hacks that lets
them work very early, and very reliably even when much of the kernel
is borked, but aren't the sort of thing you want to do except during
bringup. e.g. being mapped by special bolted TLB entries established
very early, rather than using ioremap(); polled access for interrupt
capable devices; hard-coded device addresses, meaning the kernel will
only work on one precise type of system.
You are describing the CONFIG_PPC_DEBUG_EARLY udbg methods, but they
are far from all of the available methods. Several methods use the
standard kernel apis, or do not need them (rtas, vterm). If we had
this, would we need hvc_rtas?
Stated differently, if your routine (1) fundamently works one
character
at a time and (2) is not interrupt driven, and (3) only supports one
channel, what avantage is there to an explicit hvc driver?
Using standard kernel interfaces, instead of random early debug hacks.
I was thinking of the less hackish udbg hooks that registsered in the
platform setup timeframe. Not requiring a pinned tlb that you would
not want for io performance anyways I can see.
Actually it looks remarkably similar to a cleaned up version of a
patch
i've been using since hvc_console was split to be a hookable shell.
Or
was it the motivation for adding the hooks?
Heh, ok.
...
Um.. well.. if you have both activated, I think you can select which
HVC console backend will be used by using console=/dev/hvcNN on the
commandline, where values of NN correspond to different backends, in
order depending on link order in some complex fashion.
It doesn't just work that way because of the way the hvc_console shell
implements the console hooks.
If the drivers request different hvc channel numbers, then yes, you
can
select which hvc backend is used by choosing console=hvcNN. If all
request 0, then which ever one registers with hvc_console first wins
the
slot, and the others get a busy error return. They can still
register
their tty later, at which point they get a dynamically determined
/dev/hvcN slot, above all registered console slots.
Ah, ok, my mistake.
There are multiple reasons for this design. The original user,
pseries
vterm, wants to assign the channel number based on what the hypervisor
tells it the channel should be. There are two loops though the device
tree, searching for different protocols (and consequently different
backends), so they need to be able to specify the number and not just
use the order they are found. The platform describes not only the
protocol, but which instance that is, and the backend currently
registers with that number. Another reason is its a bad idea to have
console=/dev/hvc0 depend on link order instead of knowing which driver
is selected. If a user is specifing what the console is, it should
not
depend on other linked drivers.
Until this point, all mainline drivers have been exclusive in that
only
one will actually register with hvc_console midlayer on any given
platform. This is the first backend that is not exclusive, and
therefore its coexistence needs special attention.
In my internal tree, I register (my version of) this not as hvc0 but
as
hvc4. hvc0 is rtas, hvc1-3 are an internal backend, and this one is
hvc4. The order I chose is arbitrary, but the main point is it does't
compete with hvc_rtas for slot 0. (All of these drivers coexist, and
I
can choose which one I want to use based on my needs for that boot.)
We
could make it configurable via Kconfig, or just choose a slot. (I
think
vterm can have 2 channels on some boxes where they drive real serial
ports on the box.)
Hrm, ok. Not sure what the sensible way to do this for my case would
be.
But if you already have a working HVC console driver, I don't see why
you'd want to use HVC_UDBG - it's essentially a bringup hack.
As I said, why should we need a fancier hvc_console backend if you are
polling and gain no efficency processing multiple characters at once.
The udbg drivers know if they are sharing the interface with a
hvc_console some other driver. If we trigger the registration of this
backend on the udbg hook saying its useful (eg by setting a flag word
in
the function where we assign the udbg putc and getc pointers) then
this
driver will not need to be compiled out for "production" kernels.
David Gibson wrote at 2008-11-19 00:42:20:
On Tue, Nov 18, 2008 at 09:06:17AM -0600, Timur Tabi wrote:
David Gibson wrote: [at 2008-11-18 05:14:08]
Given the variety of strange I/O configurations in prototype and
embedded platforms, I can't imagine this was a unique situation.
So
I've pushed my patch out, so anyone else in a similar situation can
immediately turn their little udbg methods for whatever strange I/O
they have into a fully-functional console. Maybe it's not
something
you'd want to go to release with, but it certainly simplifies life
during bringup.
Ok, I understand now. However, I would like to see two changes:
1) Re-arrange the Makefile as I pointed out in another post.
[the other post]
Timur Tabi - 2008-11-17 20:18:31
One other thing ...
HVC console drivers must be compiled before hvc_console.o
see http://git.kernel.org/?p=linux/kernel/git/paulus/powerpc.git;
a=commit;h=938473b24636d77dc5e9c3f41090d071b6cf4389
[end other post]
Um.. yeah.. I'm a bit baffled by this.. all the existing backends
are listed after hvc_console, I just added hvc_udbg to the end. I
didn't really understand the rationale in that commit, but then I
haven't had time to look at it very much yet.
I too challanged the necessity of that change. I carefully designed
the
hvc_console layer to find the selected console weither it was
registered
before or after the console_initcall of hvc_console.
I am very disapointed that the changelog is so sparse.
Uh... but you're the author of that commit...?
Hmm, it would appear I was the author of the patch. I'm still
disapointed, maybe my standards are higher after a few years. I was
thinking that was externally triggered.
However, I think the bottom line was that add_preferred_console is
suppsoed to be called by architecture setup code. Doing it the
console_initcall is almost abuse. But if the console_initcall is
going
to call add_preferred_console, then it must link before the
hvc_console
driver. I would have to go back and find the discussion to remember
the exact details. (maybe it only needed to call
add_preferreed_console
before registering itself if its the only hvc backend).
Which does bring up the point of avoinding calling
add_preferred_console
to udbg just because it is linked in. The udbg getc routines might
work
for xmon, but one would probably not like the result of the real
driver
and udbg trying to read the same device at the same time. Perhaps the
same flag that says "consider me or udbg_hvc" should apply, or we do
add_preferred_console in arch code before console_initcall. (its ok
to
preferr a console that never registers as long as another preferred
console does register).
Ah, yeah. I'm not sure if my hvc_udbg really wants an
add_preferred_console. It's there because it was in hvc_rtas or
whichever other hvc backend I based this on (forgotten which, now).
And the semantics of console selection makes my brain hurt, so I'm not
sure if it's actually needed for this case or not.
I think part of the problem is dummy_console ends up as a default too
often. I'm not quite sure what its purpose really is. I'm not
actually sure it is the problem either.
2) Update the Kconfig help file to be very clear that this feature
is only meaningful if the platform has a udbg back-end but no other
console or TTY driver.
Alrighty...
That does bring up the point, I don't see any kconfig help for this
user-selectable option.
Yeah, that's a good point.
And if we don't have the udbg hooks say to enable it, then we really
need to have text discouraging its use as it might break their
console.
Possibly it should just be dependent on CONFIG_PPC_EARLY_DEBUG.
If you do that then I will have to patch it back out. I choose which
hvc based on
I think that is too restrictive. I use udbg as one of my hvc
channels, but I don't use PPC_EARLY_DEBUG, that requires too much
platform knowledge. A simple flag that says 0 = only udbg-console, 1 =
allow hvc_udbg, 2 = prefer hvc_udbg that is set with the udbg hooks
would suffice for me.
That's all I have time to write at the moment, I might reply to more
later.
milton
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@ozlabs.org
https://ozlabs.org/mailman/listinfo/linuxppc-dev