Mapping sk_buff data to userspace.

I have a need to fetch sk_buff-based packets from my Loadable Kernel Module that utilizes a Netfilter hook and to present them to user space in an efficient manner.

The best overall approach I have come up with is:

Userspace calls my LKM once to initialize things:

LKM creates a 64KB VMA on behalf of the user proces using:

     ulVaddr = do_mmap_pgoff(NULL,     /* file        */
                                               0L,      /* Address: 0 = pick VMA for us. */
                                               ulLength,    /* Length       */
                                               PROT_READ | PROT_WRITE, /* Protocol.      */
                                               MAP_SHARED,    /* Flags.   */
                                               0);      /* Offset       */
       if (((long)ulVaddr) < 0)
    {
              printk(KERN_ALERT"do_mmap_pgoff error = %lu.\n", ulVaddr);
              return(NULL);
    }
/* At this point we have a valid VMA added to the calling processes address space. There is a */
/* new VMA for the returned address. Now, we find it and then muck with it to map it to our   */
/* kernel allocated memory using kmalloc(). Note, to do so with vmalloc requires additional   */
/* stuff. See notes.                    */

/* Step two: Fetch the VMA that was just created for the virtual address. */
/* We save this for later access to the user space. */
pVma = find_vma(current->mm, ulVaddr);

Then for each packet netfilter sends to my LKM, we queue it and then require the user process to call a GetNextPacket() call which does:

pvKernel = pNextSkb->head; // Start of IP Header.

iSize = PAGE_ALIGN(pNextSkb->end - pNextSkb->head);

for (ulVaddr=(unsigned long)pvKernel; ulVaddr<(unsigned long)pvKernel + iSize;ulVaddr+=PAGE_SIZE)
{
  /* Reserve all pages to make them remapable. remap_page_range() will ignore mapping if this is not done. */
  SetPageReserved(virt_to_page(ulVaddr));
}
/* Now, map the kernel memory into the user's VMA. */
/* Set mm semaphore per memory.c. */
down(&current->mm->mmap_sem);
/* We had better hit this with a previous page(s) PG_reserved set or else () frees the mapped pages and later */
/* when kfree() runs the kernel panics and the system stops(if we are VERY lucky). Meaning, it can do far worse */
/* things!    */
if (remap_page_range(pVma->vm_start,
      virt_to_phys(pvKernel),
      iSize,
      PAGE_SHARED))
{
  up(&current->mm->mmap_sem);
  printk(KERN_ALERT"remap_page_range() error for K=%p, V=%08lX, Size=%d.\n",
    pvKernel,
    pVma->vm_start,
    iSize);
  return(-EAGAIN);
}
up(&current->mm->mmap_sem);

This approach works just fine except for one thing, a bunch of "forget_pte: old mapping existed!" messages in syslog. I traced to mm/memory.c#287. This message is printed since I am re-mapping a Page Table Entry that is already mapped. The call then frees the PTE and all appears well. There is a function zap_page_range() called by mem_unmap() which does what I need but it is not exported as a Kernel symbol. When I commented out printk() in forget_pte() in mm/memory.c#287 and rebuilt 2.4.1 kernel, my code seems to work over hundreds of millions of frames with no apparent problems.

I would eventually like to release a product based on the technique above but am limited by the printk() in mm/memory.c#287.

Any suggestions on how to proceed?

Thanks in advance.

Stuart