Skip to contentSkip to conversation

virtex_devices.c

DISCUSSION Grant Likely 2 participants 5 messages May 5, 2007 – May 7, 2007 1 message not in the archive

lore
Share & tools
lore
#1 lore raw
In reply to a message not in the archive
Quoted 19 lines · David H. Lynch Jr.

I am in the process of migrating my BSP code - which is loosely
based on your earlier xilinx_mlxxx code to your newer code.
I am pulling your code from your virtex-dev branch of your git tree.

The first big obstacle I have encountered is the assumption that any
early serial device will be an 8250.
When I altered your earlier code to deal with other early serial
devices (my BSP supports both the UartLite and a pseudo serial device used when the Pico card is hosted called the Keyhole).

I made 2 significant changes:
I completely eliminated the use of plat_serial8250 in the
platform data, and used uart_port as the structure for platform data.
It had everything I needed for 8250's and could be messaged for most
anything else I had to deal with.
I explicitly set the type field in the structure so that I could use
it do decide the approriate init routine to call.

Quoted 3 lines

I do not like this - as adding additional potential early serial
devices means creating an ever increasing set of tests.
But it was better that only supporting the 8250.

I think I understand what you're describing, but I'd need to see your code snippits to be sure. Can you send me the relevant patches?

Quoted 8 lines

At some point if I get inspired I may take a wack at crafting
generic approach using an early serial mini driver.

But for the moment my choices are:
add a few more 8250 specific #ifdefs to kill off most of the
conflicting code in virtex_devices.c
and adapt my own code that is already in my BSP. This is the
simplest and does the least damage to your code.

Considering the fact that the whole purpose of virtex_devices.c is to massage the mess of #defines that is xparameters.h in to something more parsable, then extra #if defined() test may not be a big deal.

However, I need clarification. Are you talking about early serial port support, console support or regular serial devices? For example; on one of my boards here I've got both 16550 and uartlite devices on the same board and it works fine for serial port access. I've also got another design with only uartlite and another with only 16550. Each of these scenarios should work with the code that is now in mainline (but there are some issues still).

plat_serial8250_port is used because it is the structure used by the platform bus to connect the 8250 driver to 8250 devices. This is of course specific to the 8250 driver. The Uartlite ports are simple enough that they don't need a *_platform_data structure; instead the resources table in a virtex_platform_devices[] entry is sufficient to describe the device to the platform bus. You should be able to do the same thing with your keyhole device registration.

Now, early serial is still a problem. When using zImages, you must make sure that only one of CONFIG_SERIAL_UARTLITE_CONSOLE and CONFIG_SERIAL_8250_CONSOLE is set. If they are both set, then it will not compile. (But this is just a zImage boot wrapper issue. The kernel proper should work fine). Also, early serial in the kernel proper is not implemented in driver which is in mainline, but console and regular serial access is supported.

To add early serial support, I think virtex_early_serial_map() should be reworked to scan the whole virtex_platform_devices table looking for either "uartlite", "serial8250" or "keyhole" devices.

Quoted 5 lines

Or restructure your code in virtex_devices so that it uses
uart_port as the pdata format and has support to
initialize devices other than the 8250. Which I think is the better
way to go, but before I go stomping all over your
serial code I thought I was ask your thoughts ?

Go ahead, stomp away. Worst that can happen is that I disagree. :) I certainly won't be offended.

Quoted 4 lines

Is there some compelling reason to use plat_serial8250 as
the structure in which the port data is stored as opposed to
something more generic ?

plat_serial8250 isn't supposed to be generic. The generic bit is supposed to be the virtex_platform_devices table. BTW, my personal opinion is that plat_serial8250 is poorly implemented because it maps multiple 8250 devices to a single platform_device. Instead, it should be a 1-1 mapping (like all the other devices in the virtex_platform_devices table). ie. Don't try to describe all serial devices, regardless of type, in a single platform_device record.

Quoted 2 lines

Do you have an objection to my making the changes needed to
switch to the uart_port struct as the initialization data ?

It doesn't smell right; but I'm not sure I fully understand what you're suggesting. Show me patches.

Quoted 3 lines

Do you have any objection to my adding additional choices
such as UartLite to the virtex_devices serial initialization ?

Absolutely not. This is a good thing.

Quoted 4 lines

Do you have a different direction you would prefer to see things go
- aside from moving to device trees and/or implimenting
a more generic early serial driver, both of which I don't think I
want to try at this instant ?

As I described above.

Now, all this being said; we do need to migrate to arch/powerpc relatively soon. If the changes are not too complex or if they ease our migration to arch/powerpc, then I'm all for it. If you're undertaking a major development effort of stuff that's arch/ppc only; then I recommend against it. :-)

Oh, and I'd *really* like to see pico support hit mainline. Easiest way to do this is to break of chunks and tackle them one at a time which is easier than trying to get the whole patchset in at once. Keyhole driver sounds like a good place to start.

Cheers,
g.

Signature · 4 lines
Grant Likely, B.Sc. P.Eng.
Secret Lab Technologies Ltd.
[email protected]
(403) 399-0195
#2 lore raw
Quoted 4 lines · Grant Likely

I think I understand what you're describing, but I'd need to see your code snippits to be sure. Can you send me the relevant patches?

As soon as I have something working I will post it.

Quoted 10 lines
Quoted 9 lines

At some point if I get inspired I may take a wack at crafting
generic approach using an early serial mini driver.

But for the moment my choices are:
add a few more 8250 specific #ifdefs to kill off most of
the
conflicting code in virtex_devices.c
and adapt my own code that is already in my BSP. This is the
simplest and does the least damage to your code.

Quoted 3 lines

Considering the fact that the whole purpose of virtex_devices.c is to massage the mess of #defines that is xparameters.h in to something more parsable, then extra #if defined() test may not be a big deal.

This is not just about virtex_devices. Supporting early Output on
something other than an 8250
requires #ifdef's all over creation.
This is just screaming for a more generic solution. But that needs
to happen without
much increase in complexity

Quoted 3 lines

However, I need clarification. Are you talking about early serial port support, console support or regular serial devices?

While I am specifically dealing with one area. More broadly I am
talking about
all Output prior to the real serial driver coming up. This seems to be the pre-linux stuff in arch/ppc/boot/simple. You just added a uartlite_tty.c there. I have
had one for a long time.
but that is pretty trivial.
and the slightly more complex arch/ppc/syslib/uartlite_dbg that is used by the progress stuff and
for a bit prior to the early serial
driver code.
I do not think you have code for this - but I do.

These bits are only critical when something early goes completely to
crap.
But then they are important.

Quoted 6 lines

For example;
on one of my boards here I've got both 16550 and uartlite devices on the same board and it works fine for serial port access. I've also got another design with only uartlite and another with only 16550. Each of these scenarios should work with the code that is now in mainline (but there are some issues still).

This is also relevant to me because I have another pseudo serial
device that I support
in exactly the same way as UartLite. It is unique enough to our
hardware that
it is likely of little interest to anyone else. But it means
everytime I look at a UartLite
issue, I am looking at a "keyhole" issue too.
It means that everywhere there is an 8250 specific solution to a
problem,
I end up with a 3 way fork.
While support for early non-8250's is lite - there are others
besides the uartlite and my keyhole.

I rarely use JTAG and other hardware debugging tools. It is critical
for me to be able to
output information at most every part of the kernel load process.

Quoted 8 lines

plat_serial8250_port is used because it is the structure used by the platform bus to connect the 8250 driver to 8250 devices. This is of course specific to the 8250 driver. The Uartlite ports are simple enough that they don't need a *_platform_data structure; instead the resources table in a virtex_platform_devices[] entry is sufficient to describe the device to the platform bus. You should be able to do the same thing with your keyhole device registration.

By substituting uart_port for plat_serial8250 I end up with
something generic.
I do not understand the rational behind the plat_serialxxxx structs. It is not like u-boot or something else passes them, they seem to be
entirely a kernel
creation, they contain no information that is not in uart_port which
is generic for
anything that pretends to be a serial device, and using the
plat_serial8250 means
individually copying values from that struct to a uart_port. By just substitution uart_port for plat_serial8250 I end up with
code that is
exactly the same for anything that looks like a serial device.

Quoted 9 lines

Now, early serial is still a problem. When using zImages, you must make sure that only one of CONFIG_SERIAL_UARTLITE_CONSOLE and CONFIG_SERIAL_8250_CONSOLE is set. If they are both set, then it will not compile. (But this is just a zImage boot wrapper issue. The kernel proper should work fine). Also, early serial in the kernel proper is not implemented in driver which is in mainline, but console and regular serial access is supported.

Quoted 3 lines

To add early serial support, I think virtex_early_serial_map() should be reworked to scan the whole virtex_platform_devices table looking for either "uartlite", "serial8250" or "keyhole" devices.

That is basically what I am doing.
The UART macro(s) in xparameters, are changed to use uart_port
element names.
All serial/pseudo serial devices are setup uniformly the same. Serial devices are distinguishable in the uart_port struct by
their .type field,
In the rare instances I need to distinguish there are if's or
switches on the .type field.

Quoted 17 lines

Go ahead, stomp away. Worst that can happen is that I disagree. :) I certainly won't be offended.

Quoted 4 lines

Is there some compelling reason to use plat_serial8250 as
the structure in which the port data is stored as opposed to
something more generic ?

plat_serial8250 isn't supposed to be generic. The generic bit is supposed to be the virtex_platform_devices table. BTW, my personal opinion is that plat_serial8250 is poorly implemented because it maps multiple 8250 devices to a single platform_device. Instead, it should be a 1-1 mapping (like all the other devices in the virtex_platform_devices table). ie. Don't try to describe all serial devices, regardless of type, in a single platform_device record.

Quoted 5 lines
Quoted 2 lines

Do you have an objection to my making the changes needed to
switch to the uart_port struct as the initialization data ?

It doesn't smell right; but I'm not sure I fully understand what you're suggesting. Show me patches.

The basic fundimental question is why is there an 8250 specific
structure,
when there is a perfectly good generic structure that already exists ? Actually it seems to be even more complex than that because I think
there is actually an
old_serial structure that is also used in places.

This is not really a virtex or ppc question and probably belongs on
linux-serial,
but the gist is uart_port is what the majority if not all serial
devices end up using.
instead of starting with something else and then having incongruent
field names
lots fo device specific code and/or #ifdef's,
if we just use uart_port instead of plat_serialxxx's for every device, then all code gets simpler as does adding additional devices.

Shortly I hope. I have code, it is just not working at this minute. Actually I suspect it is working, but my overall migration from my implimentation of your earlier code to my implimentation of your newer code is not yet working.

Quoted 29 lines
Quoted 4 lines

Do you have any objection to my adding additional choices
such as UartLite to the virtex_devices serial initialization ?

Absolutely not. This is a good thing.

Quoted 4 lines

Do you have a different direction you would prefer to see things go
- aside from moving to device trees and/or implimenting
a more generic early serial driver, both of which I don't think I
want to try at this instant ?

As I described above.

Now, all this being said; we do need to migrate to arch/powerpc relatively soon. If the changes are not too complex or if they ease our migration to arch/powerpc, then I'm all for it. If you're undertaking a major development effort of stuff that's arch/ppc only; then I recommend against it. :-)

Oh, and I'd *really* like to see pico support hit mainline. Easiest way to do this is to break of chunks and tackle them one at a time which is easier than trying to get the whole patchset in at once. Keyhole driver sounds like a good place to start.

Cheers,
g.

Signature · 8 lines
Dave Lynch 					  	    DLA Systems
Software Development:  				         Embedded Linux
717.627.3770 	       [email protected] 	  http://www.dlasys.net
fax: 1.253.369.9244 			           Cell: 1.717.587.7774
Over 25 years' experience in platforms, languages, and technologies too numerous to list.

"Any intelligent fool can make things bigger and more complex... It takes a touch of genius - and a lot of courage to move in the opposite direction."
Albert Einstein
#3 lore raw
Quoted 13 lines · David H. Lynch Jr.
Quoted 10 lines
Quoted 9 lines

At some point if I get inspired I may take a wack at crafting
generic approach using an early serial mini driver.

But for the moment my choices are:
add a few more 8250 specific #ifdefs to kill off most of
the
conflicting code in virtex_devices.c
and adapt my own code that is already in my BSP. This is the
simplest and does the least damage to your code.

Quoted 3 lines

Considering the fact that the whole purpose of virtex_devices.c is to massage the mess of #defines that is xparameters.h in to something more parsable, then extra #if defined() test may not be a big deal.

Quoted 6 lines

This is not just about virtex_devices. Supporting early Output on
something other than an 8250
requires #ifdef's all over creation.
This is just screaming for a more generic solution. But that needs
to happen without
much increase in complexity

Okay if I understand you correctly, you're talking about two areas:
1. setting up output in the zImage wrapper (which is easy), and
2. Stuff turned on by CONFIG_SERIAL_TEXT_DEBUG which does seem to be a mess of #ifdefs.

Uartlite s already supported in the zImage wrapper, and keyhole support should be easy to add. As far as CONFIG_SERIAL_TEXT_DEBUG support; it's probably a case of do whatever you need to make it work; but don't get too worked up about getting it into mainline. It's all pretty arch/ppc specific at the moment, and if you want to work on a generic solution then you should probably wait until arch/powerpc is working.

Quoted 14 lines
Quoted 2 lines

However, I need clarification. Are you talking about early serial port support, console support or regular serial devices?

While I am specifically dealing with one area. More broadly I am
talking about
all Output prior to the real serial driver coming up. This seems to be the pre-linux stuff in arch/ppc/boot/simple. You just added a uartlite_tty.c there. I have
had one for a long time.
but that is pretty trivial.
and the slightly more complex arch/ppc/syslib/uartlite_dbg that is used by the progress stuff and
for a bit prior to the early serial
driver code.
I do not think you have code for this - but I do.

No, I've not tackled this, and I haven't looked at your code.

Quoted 10 lines

These bits are only critical when something early goes completely to
crap.
But then they are important.

Quoted 6 lines

For example;
on one of my boards here I've got both 16550 and uartlite devices on the same board and it works fine for serial port access. I've also got another design with only uartlite and another with only 16550. Each of these scenarios should work with the code that is now in mainline (but there are some issues still).

Quoted 4 lines

This is also relevant to me because I have another pseudo serial
device that I support in exactly the same way as UartLite. It is unique enough to our hardware that it is likely of little interest to anyone else. But it means everytime I look at a UartLite

Bullshit. :-) The pico is a commercially available board; support for it should be in mainline. Besides, support in mainline should make it easier for Pico to sell boards. :-)

Quoted 4 lines

issue, I am looking at a "keyhole" issue too. It means that everywhere
there is an 8250 specific solution to a problem, I end up with a 3 way fork. While support for early non-8250's is lite - there are others besides the uartlite and my keyhole.

Right; I agree this should be fixed. However keep in mind that if you work on it in arch/ppc; you'll probably just end up redoing all your work again in arch/powerpc.

Quoted 7 lines
Quoted 7 lines

plat_serial8250_port is used because it is the structure used by the platform bus to connect the 8250 driver to 8250 devices. This is of course specific to the 8250 driver. The Uartlite ports are simple enough that they don't need a *_platform_data structure; instead the resources table in a virtex_platform_devices[] entry is sufficient to describe the device to the platform bus. You should be able to do the same thing with your keyhole device registration.

Quoted 12 lines

By substituting uart_port for plat_serial8250 I end up with something generic. I do not understand the rational behind the plat_serialxxxx structs. It is not like u-boot or something else passes them, they seem to be
entirely a kernel
creation, they contain no information that is not in uart_port which
is generic for
anything that pretends to be a serial device, and using the
plat_serial8250 means
individually copying values from that struct to a uart_port.
By just substitution uart_port for plat_serial8250 I end up with
code that is
exactly the same for anything that looks like a serial device.

plat_serial8250 isn't supposed to be generic. You shouldn't use it as such. It is supposed to be the driver specific data for the driver's platform device binding. I may not like the approach of plat_serial8250, but it at the correct level, and therefore isolated. The 8250 platform bus binding is the only user of that data.

The generic bit is the platform_device structure. the virtex_platform_devices[] table is supposed to contain one entry per device. (8250 is an abomination in this regard; it uses 1 platform_device entry to represent all 8250 serial ports). The generic early serial code should *not* be looking to the virtex_serial_platform_data[] table to find ports. It should be parsing the virtex_platform_devices[] table.

It is done this way because the platform bus infrastructure takes care of binding platform_device records to platform bus device drivers. No extra code required.

Quoted 5 lines
Quoted 4 lines

Now, early serial is still a problem. When using zImages, you must To add early serial support, I think virtex_early_serial_map() should be reworked to scan the whole virtex_platform_devices table looking for either "uartlite", "serial8250" or "keyhole" devices.

Quoted 8 lines

That is basically what I am doing.
The UART macro(s) in xparameters, are changed to use uart_port
element names.
All serial/pseudo serial devices are setup uniformly the same. Serial devices are distinguishable in the uart_port struct by
their .type field,
In the rare instances I need to distinguish there are if's or
switches on the .type field.

Yes, but then you need to write new code to pass each port off to the correct driver; completely ignoring the platform bus code which *does the exact same thing*.

Quoted 9 lines
Quoted 2 lines

It doesn't smell right; but I'm not sure I fully understand what you're suggesting. Show me patches.

The basic fundimental question is why is there an 8250 specific
structure,
when there is a perfectly good generic structure that already exists ? Actually it seems to be even more complex than that because I think
there is actually an
old_serial structure that is also used in places.

I ask the opposite question; why try to make the 8250 structure generic when there is already a generic structure that does exactly what you want one level up?

I *do* think that 8250 should be changed; but not in the way you're describing.

Quoted 5 lines

This is not really a virtex or ppc question and probably belongs on
linux-serial,
but the gist is uart_port is what the majority if not all serial
devices end up using.

Right; but we're not talking about the internal use of uart_port by the drivers. We're talking about the

Quoted 9 lines

instead of starting with something else and then having incongruent
field names
lots fo device specific code and/or #ifdef's,
if we just use uart_port instead of plat_serialxxx's for every device, then all code gets simpler as does adding additional devices. Shortly I hope. I have code, it is just not working at this minute. Actually I suspect it is working, but my overall migration from my implimentation of your earlier code to my implimentation of your newer code is not yet working.

I look forward to seeing it.

Cheers,
g.

Signature · 4 lines
Grant Likely, B.Sc. P.Eng.
Secret Lab Technologies Ltd.
[email protected]
(403) 399-0195
#4 lore raw
Quoted 8 lines · Grant Likely
Quoted 6 lines · David H. Lynch Jr.

The basic fundimental question is why is there an 8250 specific
structure,
when there is a perfectly good generic structure that already exists ? Actually it seems to be even more complex than that because I think
there is actually an
old_serial structure that is also used in places.

Quoted 5 lines

I ask the opposite question; why try to make the 8250 structure generic when there is already a generic structure that does exactly what you want one level up?

I *do* think that 8250 should be changed; but not in the way you're describing.

Although this point is actually moot because platform_bus pretty much goes away when we move to arch/powerpc. We'll be using of_platform_devices instead (which map onto nodes in the device tree).

Cheers,
g.

Signature · 4 lines
Grant Likely, B.Sc. P.Eng.
Secret Lab Technologies Ltd.
[email protected]
(403) 399-0195
#5 lore raw
Quoted 4 lines · Grant Likely

Although this point is actually moot because platform_bus pretty much goes away when we move to arch/powerpc. We'll be using of_platform_devices instead (which map onto nodes in the device tree).

That I got and as I pound away at all of this I keep thinking my
time would be better spent figuring out device trees.

Quoted 1 line
Signature · 8 lines
Dave Lynch 					  	    DLA Systems
Software Development:  				         Embedded Linux
717.627.3770 	       [email protected] 	  http://www.dlasys.net
fax: 1.253.369.9244 			           Cell: 1.717.587.7774
Over 25 years' experience in platforms, languages, and technologies too numerous to list.

"Any intelligent fool can make things bigger and more complex... It takes a touch of genius - and a lot of courage to move in the opposite direction."
Albert Einstein