Kernel malloc buffers how to make contiguous in user space?

[Dave Cogley]

Hello,

 

I am using a scatter gather DMA operation provided on the PPC44EPx to handle
a large data move of 2MB from sixteen 128k kmalloc buffers.  I have sixteen
scatter / gather descriptors pointing to the allocated memory which will be
transferred to a peripheral on the EBC.  This all appears to be setup and
working correctly.  Now I want to provide these 16 kmalloc buffers to a user
space process and make them all appear as a contiguous 2MB buffer.

 

When the handler is called pageptr appears to be a valid page but the kernel
traps with "Bad page state" "Trying to fix it up, but a reboot is needed"
when returned from the handler.  Am I missing something required to modify
the page state for the requested page?  Am I even handling the nopage
request correctly?  I notice the nopage handler gets called for every 4096
byte page it is trying to map.  How do I map my 128k memory area into the 4k
page requests?  Is there example code somewhere that demonstrates using a
contiguous memory buffer in user space when it is comprised of multiple
buffers in the kernel space?

 

 

First I am allocating 16 buffers using kmalloc:

 

#define DMABLOCKSIZE          1024 * 128

 

for (I = 0; I < 16; i++)

{

            device->dma_buf[i] = kmalloc(DMABLOCKSIZE, GFP_DMA |
GFP_KERNEL);

            sgl->phyaddress = virt_to_phys(device->dma_buf[i]);

}

 

Then I am doing the following in the nopage handler:

 

page* nopage_handler(struct vma, ul address, int* type)

{

            unsigned long physaddr = address - vma->start;

            int index = physaddr / DMABLOCKSIZE;

            int off = physaddr % DMABLOCKSIZE;

            struct page* pageptr = virt_to_page(device->dma_buf[index] +
off);

 

            get_page(pageptr);

            if (type)

                        *type = VM_FAULT_MINOR;

 

            return pageptr;

}

 

 

Dave Cogley

Software Engineer

Ultra Stereo Labs, Inc.

(805) 549-0161

mailto:dcogley at uslinc.com

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://ozlabs.org/pipermail/linuxppc-embedded/attachments/20071101/3979bd8b/attachment-0001.htm