I have an IPMI interface driver that sits on top of the I2C code. I'd
like to get it into the mainstream kernel, but I have a few problems
to solve first before I can do that. The I2C code is synchronous and
must run from a task context. The IPMI driver has certain
operations that occur at panic time, including:
* Storing panic information in IPMI's system event log
* Extending the watchdog timer so it doesn't go off during panic
operations (like kernel coredumps).
* Powering the system off
I can't really put the IPMI SMB interface into the kernel until I can
do those operations. Also, I understand that some vendors put RTC
chips onto the I2C bus and this must be accessed outside task context,
too. I would really like add asynchronous interface to the I2C bus
drivers. I propose:
* Adding an async send interface to the busses that does a callback
when the operation is complete.
* Adding a poll interface to the busses. The I2C core code could
call this if a synchronous call is made from task context (much
like all the current drivers do right now). For asyncronous
operation, the I2C core code would call it from a timer
interrupt. If the driver supported interrupts, polling from the
timer interrupt would not be necessary.
* Add async operations for the user to call, including access to the
polling code.
* If the driver didn't support an async send, it would work as it
does today and the async calls would return ENOSYS.
This way, the bus drivers on I2C could be converted on a
driver-by-driver basis. The IPMI code could query to see if the
driver supported async operations. And the RTC code could use it,
too.
Is this ok with the I2C community? I would do the base work and
convert over a few drivers.
Thanks,
-Corey
Quoth Corey Minyard:
> I would really like add asynchronous interface to the I2C bus drivers.
Applause! This is IMO overdue, but maybe sensor systems don't need it
as much as other I2C applications do. For example, see the isp1301_omap
driver, which could have been hugely simpler if there were an async I2C
framework to build on. That's probably one of the more complex examples
floating around ... but it's hardly the only place where needing to talk
synchronously to I2C chips creates trouble.
> I propose:
>
> * Adding an async send interface to the busses that does a callback
> when the operation is complete.
It'd have to have a queue, so that several different chips could have
operations pending concurrently. I suspect you wouldn't need to cancel
operations once they're queued ... useful simplification, compared to
for example USB. (Which in retrospect needs a "kill queue" operation,
per endpoint, rather than the "cancel request" operation we've got.
But when 2.2 started with USB, we didn't quite know that ... ;)
> * Adding a poll interface to the busses. The I2C core code could
> call this if a synchronous call is made from task context (much
> like all the current drivers do right now). For asyncronous
> operation, the I2C core code would call it from a timer
> interrupt. If the driver supported interrupts, polling from the
> timer interrupt would not be necessary.
> * Add async operations for the user to call, including access to the
> polling code.
> * If the driver didn't support an async send, it would work as it
> does today and the async calls would return ENOSYS.
To the extent I've thought about it, that sounds like a good approach.
- Dave
is there a way to do this solely in i2c-core without having to
add support to all the drivers?
Corey Minyard wrote:
> I have an IPMI interface driver that sits on top of the I2C code. I'd
> like to get it into the mainstream kernel, but I have a few problems
> to solve first before I can do that. The I2C code is synchronous and
> must run from a task context. The IPMI driver has certain
> operations that occur at panic time, including:
>
> * Storing panic information in IPMI's system event log
> * Extending the watchdog timer so it doesn't go off during panic
> operations (like kernel coredumps).
> * Powering the system off
>
> I can't really put the IPMI SMB interface into the kernel until I can
> do those operations. Also, I understand that some vendors put RTC
> chips onto the I2C bus and this must be accessed outside task context,
> too. I would really like add asynchronous interface to the I2C bus
> drivers. I propose:
>
> * Adding an async send interface to the busses that does a callback
> when the operation is complete.
> * Adding a poll interface to the busses. The I2C core code could
> call this if a synchronous call is made from task context (much
> like all the current drivers do right now). For asyncronous
> operation, the I2C core code would call it from a timer
> interrupt. If the driver supported interrupts, polling from the
> timer interrupt would not be necessary.
> * Add async operations for the user to call, including access to the
> polling code.
> * If the driver didn't support an async send, it would work as it
> does today and the async calls would return ENOSYS.
>
> This way, the bus drivers on I2C could be converted on a
> driver-by-driver basis. The IPMI code could query to see if the
> driver supported async operations. And the RTC code could use it,
> too.
>
> Is this ok with the I2C community? I would do the base work and
> convert over a few drivers.
>
> Thanks,
>
> -Corey
>
>
Mark Studebaker wrote:
> is there a way to do this solely in i2c-core without having to
> add support to all the drivers?
Yes and no. In order to support this async operation, the driver cannot
block and do things like msleep() or schedule(). It has to start the
operation, return, and either let polling or an interrupt drive the
continued operation. Thus for async operations the driver has to be
modified. However, if async operation is not required, the driver can
stay as is.
I've been working on this and will probably have a patch tomorrow. I've
modified the piix4 and the i801 drivers, I probably won't do any more
myself unless the need arises, since I can't test any others. Note that
this still supports the old driver interface, so no drivers need to be
rewritten. That way, they only need to be modified if something needs
the async interface. So drivers that have an RTC on them or that
support IPMI BMCs could be rewritten, but nothing else needs to be done.
I've also noticed a somewhat cavalier attitude in this code with respect
to return values. I've cleaned some of that up so return values are not
just -1 on error, but are proper errno values. However, I've only fixed
the core code and the drivers I've worked on.
Thanks,
-Corey
>
> Corey Minyard wrote:
>
>> I have an IPMI interface driver that sits on top of the I2C code. I'd
>> like to get it into the mainstream kernel, but I have a few problems
>> to solve first before I can do that. The I2C code is synchronous and
>> must run from a task context. The IPMI driver has certain
>> operations that occur at panic time, including:
>>
>> * Storing panic information in IPMI's system event log
>> * Extending the watchdog timer so it doesn't go off during panic
>> operations (like kernel coredumps).
>> * Powering the system off
>>
>> I can't really put the IPMI SMB interface into the kernel until I can
>> do those operations. Also, I understand that some vendors put RTC
>> chips onto the I2C bus and this must be accessed outside task context,
>> too. I would really like add asynchronous interface to the I2C bus
>> drivers. I propose:
>>
>> * Adding an async send interface to the busses that does a callback
>> when the operation is complete.
>> * Adding a poll interface to the busses. The I2C core code could
>> call this if a synchronous call is made from task context (much
>> like all the current drivers do right now). For asyncronous
>> operation, the I2C core code would call it from a timer
>> interrupt. If the driver supported interrupts, polling from the
>> timer interrupt would not be necessary.
>> * Add async operations for the user to call, including access to the
>> polling code.
>> * If the driver didn't support an async send, it would work as it
>> does today and the async calls would return ENOSYS.
>>
>> This way, the bus drivers on I2C could be converted on a
>> driver-by-driver basis. The IPMI code could query to see if the
>> driver supported async operations. And the RTC code could use it,
>> too.
>>
>> Is this ok with the I2C community? I would do the base work and
>> convert over a few drivers.
>>
>> Thanks,
>>
>> -Corey
>>
>>
> I will be glad to work with on this, I have some exposure to the BMC. See text below in blue.
>
> bukie
>
> Corey Minyard wrote:
>
>> Mark Studebaker wrote:
>>
>> > is there a way to do this solely in i2c-core without having to
>> > add support to all the drivers?
>>
>> Yes and no. In order to support this async operation, the driver cannot
>> block and do things like msleep() or schedule().
>
> The Interface driver must be a user space driver.
>
>> It has to start the
>> operation, return, and either let polling or an interrupt drive the
>> continued operation. Thus for async operations the driver has to be
>> modified. However, if async operation is not required, the driver can
>> stay as is.
>>
>> I've been working on this and will probably have a patch tomorrow. I've
>> modified the piix4 and the i801 drivers, I probably won't do any more
>> myself unless the need arises, since I can't test any others. Note that
>> this still supports the old driver interface, so no drivers need to be
>> rewritten. That way, they only need to be modified if something needs
>> the async interface. So drivers that have an RTC on them or that
>> support IPMI BMCs could be rewritten, but nothing else needs to be done.
>>
>> I've also noticed a somewhat cavalier attitude in this code with respect
>> to return values. I've cleaned some of that up so return values are not
>> just -1 on error, but are proper errno values. However, I've only fixed
>> the core code and the drivers I've worked on.
>>
>> Thanks,
>>
>> -Corey
>>
>> >
>> > Corey Minyard wrote:
>> >
>> >> I have an IPMI interface driver that sits on top of the I2C code. I'd
>> >> like to get it into the mainstream kernel, but I have a few problems
>> >> to solve first before I can do that. The I2C code is synchronous and
>> >> must run from a task context.
>
> I am not sure what you mean that the I2C code is synchronous. I2C bus is Asynchronous which means that the data clock is not included in the data. The Sender and Receiver agrees on the timing parameters prior to the bus transaction.
>
>> The IPMI driver has certain
>> >> operations that occur at panic time, including:
>> >>
>> >> * Storing panic information in IPMI's system event log
>> >> * Extending the watchdog timer so it doesn't go off during panic
>> >> operations (like kernel coredumps).
>> >> * Powering the system off
>> >>
>
> Is this driver compliant with the IPMI spec? Because the above should be a sensor that must be enable or disable. A driver should not make sure a decision by itself.
>
>>
>> >> I can't really put the IPMI SMB interface into the kernel until I can
>> >> do those operations. Also, I understand that some vendors put RTC
>> >> chips onto the I2C bus and this must be accessed outside task context,
>> >> too.
>
> What the vendor put on the board doesn't matter with respect to IPMI. What matter is that you have access to the Master where the slave you talking to is connect on the I2C bus.
>
>> I would really like add asynchronous interface to the I2C bus
>> >> drivers.
>
> Do you mean a blocking and non blocking I/O?
>
>> I propose:
>> >>
>> >> * Adding an async send interface to the busses that does a callback
>> >> when the operation is complete.
>
> Okay, you are doing a non-blocking I/O. So what happens when another process tries to access the I2C bus before the bus transaction is completed. Some thing (mainly the app) needs to tell the driver not to share the resource while a transaction is in progress.
>
>>
>> >> * Adding a poll interface to the busses. The I2C core code could
>> >> call this if a synchronous call is made from task context (much
>> >> like all the current drivers do right now). For asyncronous
>> >> operation, the I2C core code would call it from a timer
>> >> interrupt. If the driver supported interrupts, polling from the
>> >> timer interrupt would not be necessary.
>
> I think this should be done in the Interface code because the Interface code will be running in the user space and have access to the operating system facility.
>
>>
>> >> * Add async operations for the user to call, including access to the
>> >> polling code.
>
> The driver can make itself blocking and non blocking
>
>>
>> >> * If the driver didn't support an async send, it would work as it
>> >> does today and the async calls would return ENOSYS.
>
> Not needed, it will be addressed by the blocking and non-block implementation of the driver.
>
>>
>> >>
>> >> This way, the bus drivers on I2C could be converted on a
>> >> driver-by-driver basis. The IPMI code could query to see if the
>> >> driver supported async operations. And the RTC code could use it,
>> >> too.
>> >>
>> >> Is this ok with the I2C community? I would do the base work and
>> >> convert over a few drivers.
>> >>
>> >> Thanks,
>> >>
>> >> -Corey
>> >>
>> >>
>>
>> -
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at http://www.tux.org/lkml/
>
Bukie Mabayoje wrote:
> I will be glad to work with on this, I have some exposure to the BMC.
> See text below in blue.
>
> bukie
>
> Corey Minyard wrote:
>
>> Mark Studebaker wrote:
>>
>> > is there a way to do this solely in i2c-core without having to
>> > add support to all the drivers?
>>
>> Yes and no. In order to support this async operation, the driver cannot
>> block and do things like msleep() or schedule().
>>
> The Interface driver must be a user space driver.
That is precisely the problem. What happens if you need to get to the
i2c bus but you are not running from a task context?
>> It has to start the
>> operation, return, and either let polling or an interrupt drive the
>> continued operation. Thus for async operations the driver has to be
>> modified. However, if async operation is not required, the driver can
>> stay as is.
>>
>> I've been working on this and will probably have a patch tomorrow. I've
>> modified the piix4 and the i801 drivers, I probably won't do any more
>> myself unless the need arises, since I can't test any others. Note that
>> this still supports the old driver interface, so no drivers need to be
>> rewritten. That way, they only need to be modified if something needs
>> the async interface. So drivers that have an RTC on them or that
>> support IPMI BMCs could be rewritten, but nothing else needs to be done.
>>
>> I've also noticed a somewhat cavalier attitude in this code with respect
>> to return values. I've cleaned some of that up so return values are not
>> just -1 on error, but are proper errno values. However, I've only fixed
>> the core code and the drivers I've worked on.
>>
>> Thanks,
>>
>> -Corey
>>
>> >
>> > Corey Minyard wrote:
>> >
>> >> I have an IPMI interface driver that sits on top of the I2C code.
>> I'd
>> >> like to get it into the mainstream kernel, but I have a few problems
>> >> to solve first before I can do that. The I2C code is synchronous and
>> >> must run from a task context.
>>
> I am not sure what you mean that the I2C code is synchronous. I2C bus
> is Asynchronous which means that the data clock is not included in the
> data. The Sender and Receiver agrees on the timing parameters prior to
> the bus transaction.
I mean the driver, not the hardware. By the I2C driver being
synchronous, I mean that you call a function, the whole operation occurs
in the function, and then the function returns the result. By
asynchronous, I mean that you tell the driver to start an operation and
the driver starts it and immediately returns. It then uses interrupts,
timers, or polling calls to drive the operation of the interface. When
the operation is done, the results are reported back through a callback
function provided when the driver started.
Plus there are SMB alerts which the hardware can generate
asynchronously. The I2C driver has no concept of that right now.
>> The IPMI driver has certain
>> >> operations that occur at panic time, including:
>> >>
>> >> * Storing panic information in IPMI's system event log
>> >> * Extending the watchdog timer so it doesn't go off during panic
>> >> operations (like kernel coredumps).
>> >> * Powering the system off
>> >>
>
> Is this driver compliant with the IPMI spec? Because the above should
> be a sensor that must be enable or disable. A driver should not make
> sure a decision by itself.
There are two parts here. There is the I2C driver and the IPMI driver
that sits on top of it. I'm really only taking about the I2C driver
here, I need asynchronous operation from the I2C driver to do those
things in the IPMI driver.
>>
>> >> I can't really put the IPMI SMB interface into the kernel until I can
>> >> do those operations. Also, I understand that some vendors put RTC
>> >> chips onto the I2C bus and this must be accessed outside task
>> context,
>> >> too.
>
> What the vendor put on the board doesn't matter with respect to IPMI.
> What matter is that you have access to the Master where the slave you
> talking to is connect on the I2C bus.
>
>> I would really like add asynchronous interface to the I2C bus
>> >> drivers.
>
> Do you mean a blocking and non blocking I/O?
Yes, that is a better term. I'll switch to using that.
-Corey