Re: [RFC 1/2] PCI-Express Non-Transparent Bridge Support

On Sun, Jul 15, 2012 at 05:19:21PM -0700, Greg KH wrote:

On Sun, Jul 15, 2012 at 04:55:48PM -0700, Jon Mason wrote:

quoted

On Sat, Jul 14, 2012 at 10:10:15AM -0700, Greg KH wrote:

quoted

On Fri, Jul 13, 2012 at 02:44:59PM -0700, Jon Mason wrote:

quoted

+static int max_num_cbs = 2;
+module_param(max_num_cbs, uint, 0644);
+MODULE_PARM_DESC(max_num_cbs, "Maximum number of NTB transport connections");
+
+static bool no_msix;
+module_param(no_msix, bool, 0644);
+MODULE_PARM_DESC(no_msix, "Do not allow MSI-X interrupts to be selected");

How would a user, or a distro, know to set these options?  Why are they
even options at all?

Good question.  There is actually a potential benefit to disabling
MSI-X.  The NTB device on one of our platforms only has 3 MSI-X
vectors.  In the current driver design, that would limit them to 3
client/virtual devices.  However, there are 15bits in the ISR that can
be used for the same purpose.  So, if you disable MSI-X, you can have
15 instead of 3.  

But again, how would a user, or a distro, know to set these?  Where is
the documentation describing it?  Why really have these options at all
and not just fix the platform issues (only 3 MSI-X vectors?  Really?)

I believe we'll want multiple clients (or have multiqueue Ethernet).  I'm happy to add something to /Documentation to describe it and why it would be useful, or I can remove it and re-introduce it when I add multiqueue Ethernet.

3 MSI-X vectors (plus one for PCI-E link up/down) on Xeon NTB, and 33 for Atom NTB.  Yeah, really.

thanks,

greg k-h