Question on Au1550 ...

Question on Au1550 ...

[Alexey Shinkin]

Hi , All !

I am porting a linux driver for PCI device and some applications form x86 
platform to  AMD Au1550 MIPS , running MontaVista Linux 2.4.20-mvl31.1550.

I got the  following  problem :

The driver allocates some memory (SharedMemory)  for communication with  PCI 
card , using  function pci_alloc_consistent().  PCI card can write to this 
memory using DMA .  On  kernel level  driver correctly reads values, written 
by the board , and the board  correctly reads values, written by driver.

The goal is to make this memoty visible from user level application. The 
following approach is used on õ86 and works :
1. Application calls  mmap( 0, Size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0 
) ;

2.  driver implements mmap() function  that just  prepares some vma 
structure fields for nopage operetions :
	vma->vm_flags |= VM_RESERVED;	// no swapping
	vma->vm_ops	= &device_vmops; // contains nopage func
	vma->vm_file   = filp;

3. Application  gets pointer and accesses all mapped pages (reads a byte 
from each page) - this  causes call of nopage function in driver .
The nopage function calculates VirtualAddress from SharedMemory address 
(received after pci_alloc_consistent()),
then  does   KernelPage= virt_to_page(VirtualAddress)   ,
   	get_page(KernelPage)
   and returns KernelPage.

4. Now Application  reads  from SharedMemory , using a pointer , given by 
mmap()  .
This works fine on x86 platform , but on MIPS  application cannot see 
expected contents of the memory - it reads garbage .

What I have  tried :
     - to insert   vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot)  
in driver’s implementation of mmap() ;
   -  to use  remap_page_range() instead of vm_ops .

Result is the same – application reads garbage . If application write  some 
values into the memory - it reads garbage . The values written by 
application are not seen from kernel level .


I am new to MIPS so maybe I missed  something obvious ? Is this right 
approach on this platform at all ?

Thank you in advance!

_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today it's FREE! 
http://messenger.msn.com/

Re: Question on Au1550 ...

[Ladislav Michl]

On Mon, Oct 18, 2004 at 07:01:30PM +0700, Alexey Shinkin wrote:
[snip]
> What I have  tried :
>     - to insert   vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot)
> in driver?s implementation of mmap() ;
>   -  to use  remap_page_range() instead of vm_ops .

Did you SetPageReserved bit?

	ladis

Re: Question on Au1550 ...

[Dan Malek]

On Oct 18, 2004, at 8:01 AM, Alexey Shinkin wrote:

> The goal is to make this memoty visible from user level application. 
> The following approach is used on õ86 and works :

I'm kinda surprised it works anywhere.  You have taken a rather complex
approach to achieving your results.

>   -  to use  remap_page_range() instead of vm_ops .

This is all you should have to do.  Use the physical (dma_addr_t) 
address
returned from pci_alloc_consistent() as the 'offset' parameter, use the 
VMA
allocated and provided to you, call remap_page_range() and you are done.

An alternative is to have an ioctl() in your driver that returns the 
physical
address of the buffer (as provided by pci_alloc_consistent()), then use
/dev/mem to do the mmap().  The only problem with this is coordination
with your driver, you have to ensure you release the /dev/mem mmap
in your application when the driver also does the same for the dma
buffer space.


	-- Dan