question about writing to memory directly from user space in Linux

question about writing to memory directly from user space in Linux

[Vijay Padiyar]

Hi

I am running Linux 2.6.10 on an MPC8260 target. We have an I2C controller
that is part of our application code. In VxWorks, we could address the
MPC8260 I2C memory registers directly from application space and so this was
not a problem.

To port this same code to Linux, I am opening '/dev/mem' and mmapping the
internal memory map of 128 KB to user space. I am writing directly to the
I2C_BASE, I2C PRAM and I2C RxBD/TxBD registers, all of which are in DPRAM,
from my application code through suitable macros.

All of this is fine. I can see that the values have been written to memory
if I do a memory dump of those registers with an external dump tool that
uses '/dev/mem' too.

Now, I have kept my Tx and Rx buffers in DPRAM as well, and have programmed
those addresses in the buffer pointers of my TxBD and RxBDs. However, I am
not sure if the values I'm writing to those locations from user space are
actually being written to the actual memory locations.

This is because I print the value of the first byte of my transmit buffer
pointer BEFORE and AFTER I write the I2C address (0xA0 for us) to it. Before
the address is written, the contents shown are 0x90. After it's written, the
contents are shown rightly as 0xA0, but I don't think this value is actually
written to DPRAM.

This is because when I start I2C transmission by setting the start transfer
bit in I2COM register, the address byte outputted on the I2C SDA bus is
0x90. This means it outputs the original value and not the value which I
wrote to it.

Can you suggest some kind of a "synchronization" function which I can call
from *user space* which will ensure that all values written to mmaped memory
locations from user space are actually flushed to their physical locations?
I think this might be a potential problem with my code.

Regards

Vijay Padiyar

http://www.vijaypadiyar.eu.tf

Re: question about writing to memory directly from user space in Linux

[Stefan Nickl]

On Wed, 2005-03-23 at 14:25 +0530, Vijay Padiyar wrote:
> Can you suggest some kind of a "synchronization" function which I can call
> from *user space* which will ensure that all values written to mmaped memory
> locations from user space are actually flushed to their physical locations?
> I think this might be a potential problem with my code.

Have a look at linux/include/asm-ppc/{io.h,system.h} and
http://www.xml.com/ldd/chapter/book/ch08.html

The eieio instruction doesn't need supervisor rights.

Greetings,

-- 
Stefan Nickl
Kontron Modular Computers

Re: question about writing to memory directly from user space in Linux

[Eugene Surovegin]

On Wed, Mar 23, 2005 at 10:46:45AM +0100, Stefan Nickl wrote:
> On Wed, 2005-03-23 at 14:25 +0530, Vijay Padiyar wrote:
> > Can you suggest some kind of a "synchronization" function which I can call
> > from *user space* which will ensure that all values written to mmaped memory
> > locations from user space are actually flushed to their physical locations?
> > I think this might be a potential problem with my code.
> 
> Have a look at linux/include/asm-ppc/{io.h,system.h} and
> http://www.xml.com/ldd/chapter/book/ch08.html
> 
> The eieio instruction doesn't need supervisor rights.

eieio is a nop on 603e core.

--
Eugene

Re: question about writing to memory directly from user space in Linux

[Stefan Nickl]

On Wed, 2005-03-23 at 09:10 -0800, Eugene Surovegin wrote:
> On Wed, Mar 23, 2005 at 10:46:45AM +0100, Stefan Nickl wrote:
> > On Wed, 2005-03-23 at 14:25 +0530, Vijay Padiyar wrote:
> > > Can you suggest some kind of a "synchronization" function which I can call
> > > from *user space* which will ensure that all values written to mmaped memory
> > > locations from user space are actually flushed to their physical locations?
> > > I think this might be a potential problem with my code.
> > 
> > Have a look at linux/include/asm-ppc/{io.h,system.h} and
> > http://www.xml.com/ldd/chapter/book/ch08.html
> > 
> > The eieio instruction doesn't need supervisor rights.
> 
> eieio is a nop on 603e core.

Interesting, I was not aware of this.

eieio is handled differently between MPC8245 and MPC8260.

Anyways, it's a non-issue as IO memory should have the right WIMG bits
set to take care of this, and modelling it with a in-kernel i2c bus as
suggested on the list is soo much nicer.

-- 
Stefan Nickl
Kontron Modular Computers

RE: question about writing to memory directly from user space in Linux

[Fillod Stephane]

Vijay Padiyar wrote:
>I am running Linux 2.6.10 on an MPC8260 target. We have an I2C
controller
>that is part of our application code. In VxWorks, we could address the
>MPC8260 I2C memory registers directly from application space and so
this was
>not a problem.

The I2C abstraction layer of Linux is very good, and user-land support
quite comfortable. Why bother port/develop a new driver? Besides,=20
interrupt handling in user-space can be troublesome..

>Can you suggest some kind of a "synchronization" function which I can
call
>from *user space* which will ensure that all values written to mmaped
memory
>locations from user space are actually flushed to their physical
locations?
>I think this might be a potential problem with my code.

Please have a look at Denx's FAQ for accessing memory bus:

=20
http://www.denx.de/twiki/bin/view/PPCEmbedded/DeviceDrivers#Section_Acce
ssingPeripheralsFromUserSpace
or shorter: http://tinyurl.com/6c7th

Depending on the O_SYNC flag passed to open, the mmap'ed memory will be
accessed through non guarded cache or not. See linux/drivers/char/mem.c.
The ppc instruction "eieio" deals with I/O ordering.


Regards,
--=20
Stephane