anand kumar wrote:
> We are developing a PCI driver for a specialized hardware which
> needs blocks of physically contiguous memory regions of
> 32 KB. We need to allocate 514 such blocks for a total of 16 MB
> We were using an ioctl implementation in the driver which uses
> kmalloc() to allocate the required memory blocks. 
> kmalloc()(GFP_KERNEL)
> fails after allocating some 250 blocks of memory (probably due to 
> fragmentation).
> We then tried using __get_free_pages() and the result was the 
> same.

Well, kmalloc() falls back to __get_free_pages() eventually (after going
through the slab cache), so it isn't much of a surprise that both failed.

> Even though the free pages in zone NORMAL and DMA were 10000 and
> 1500 respectively.

Can you try this on a kernel that has /proc/buddyinfo?  It exports the
buddy allocators internal structures so that you can see the
fragmentation.  buddyinfo has been in 2.5 since ~2.5.35.  There may even
be a 2.4 version floating around...

If you can post some code too, it might be helpful.
-- 
Dave Hansen
haveblue@us.ibm.com