Messages in this thread | | | From | "Pan, Weimin" <> | Date | Mon, 30 Jun 2008 20:45:47 +0000 | Subject | RE: Why get_user_pages fails? |
| |
Yes, I totally agree with you that the best solution is to bypass the Linux VM. The reason to continue using get_user_pages() is that we can leverage the existing kernel driver - the SCST.
Do I have a way to map a range of phy memory to user virtual address and have the page structures built for that memory region? I know remap_pfn_range() doesn't work.
Weimin
-----Original Message----- From: Jonathan Corbet [mailto:corbet@lwn.net] Sent: Monday, June 30, 2008 2:52 PM To: Pan, Weimin Cc: linux-kernel@vger.kernel.org Subject: Re: Why get_user_pages fails?
On Mon, 30 Jun 2008 19:32:41 +0000 "Pan, Weimin" <weimin.pan@hp.com> wrote:
> I booted the kernel and set "mem=512M" and then mmapped the physical > memory above 512M to user space and got the user space virtual address > for that memory. When I passed that address to my driver to do direct > IO. The get_user_pages() in my driver fails. I found that this is due > to the VM_IO and VM_PFNMAP are set. Does that mean the page structures > are not built when I mmapped physical memory to user space (using > remap_pfn_range())? > > Do I have a way to solve this problem? How can I map the physical > memory to user space virtual address and can pass this address to my > driver to do direct IO by using get_user_pages() function?
In this situation, you already know where the memory sits in physical space, and there's no point in trying to pin it. So a call to get_user_pages() is entirely unnecessary. You're bypassing the Linux VM, so bypass it entirely.
However, this whole scheme seems a little...baroque. Why are you doing things this way? As always in such cases, posting your code can enable others to help you find the right solution to your problem.
jon
| |