locking user memory and kiobuf

Juan Gonzalez (juang@gate.net)
Sat, 06 Nov 1999 02:38:46 -0500

I'm in the process of porting a WinNT driver to Linux. For a specific
IOCtl call, the driver receives pointers to multiple user buffers.
Based on different conditions, the driver needs to transfer data to/from
one of these buffers via DMA (actually PCI Bus Master), and then return
to the user when the transfer is complete. The user doesn't mlock any
of them, so it's up to the driver to handle it. Currently, I have this
working by first copying the user data to a driver buffer (locked) and
then DMAing from there. It works, but large buffers move pretty slow.
To improve performance, it would be nice if the driver could lock the
required user buffer and transfer directly to/from there without having
to copy. We have implemented this strategy successfully in WinNT, OS/2,
and AIX, with very good performance improvements.

Three questions:

1. From what I've read in the archives, it seems that locking user
memory from a driver isn't supported in current releases of the Linux
kernel (2.2.xx). Is that correct?

2. As I understand, there's a new "kiobuf mechanism" being added in
2.3.xx .
Does that provide a solution to this particular issue?

3. Alternatively, can I directly manipulate the PG_locked flag for the
pages in question?
For example, I could ensure the page(s) are in physical memory by
accessing a word in that page (copy_to_user), and then set the PG_locked
bit for the page to prevent swapping. Wouldn't this effectively "lock"
the page for safe DMA transfers? After the transfer, I could clear the
PG_locked bit and return to the user.
Or would this be considered dangerous practice? Do I also need
PG_reserved?

Thank you,
Juan Gonzalez

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/