Re: Zero-copy transfer with kiobufs: URGENT

[Fabrice Tchakountio <ftchakou@ftchakountio.bbn.com>]

Hi Stuart -

Thanks for your excellent suggestion. I have a couple of comments.

* How can we deal with potentially large chunks of memory:
  I thought we need vmalloc() to allocate large kernel chunks (e.g: 
>= 10 MB)

* I'm concerned about memory limitations since the data
  is still wired:

  you're suggesting a mmap-based approach with the user having the
  possibility to free up the memory with a close() call. 

  In our application, we intend to store each transferred chunk from the
  kernel in a ring buffer. Thus there is a potential risk to consume all
  the main memory with a mmap-based approach since the data is still
  wired. 

  Ideas ?

  Thanks











On Thu, 4 Oct 2001, Stuart Macdonald wrote:

> Hi,
> 
> The following works for me:
> 
> 1) User space calls kernel-based function which:
> 
>         a)  creates a Virtual Memory Area of say 64KB. do_mmap_pgoff()
> returning address and then call find_vma() to         get the actual vma.
> 
>         b) does a pvKernelMem = __get_free_pages(up to 128KB I believe) to
> allocate contiguous kernel pages. Note,         kmalloc() does not  always
> end up being page-aligned. Save pvKernelMem for kernel access to shared
> memory.
> 
>         c) Map the user VMA to the kernel pages using: set_page_reserved()
> on all pages then remap_page_range().
> 
>         d) return VMA->vm_start to the user to denote the beginning address
> of the shared memory.
> 
>  2) Use memory. Suggest you specify "volatile" on variables utilizing such
> shared memory to avoid compiler                         optimization related
> issues.
> 
> 3) Have a "Close" function unmap by using do_munmap() and free kernel pages.
> 
> 
> Stuart
> 
> 

-
Kernelnewbies: Help each other learn about the Linux kernel.
Archive:       http://mail.nl.linux.org/kernelnewbies/
IRC Channel:   irc.openprojects.net / #kernelnewbies
Web Page:      http://www.kernelnewbies.org/