In D6437#140473, @andrew wrote:

In D6437#140235, @zbb wrote:

Again, I would prefer a patch against existing ITS code not instead of it. This would allow us to track the diff properly and here it's hardly possible. Most of my previous remarks still apply ;-).

Many of the remarks are how I've done things different than the existing driver. This is expected when one of the goals was to reduce the knowledge of the internals of the GICv3 driver.

Again, I would prefer a patch against existing ITS code not instead of it. This would allow us to track the diff properly and here it's hardly possible. Most of my previous remarks still apply ;-).

I get that we have new API, new callbacks to provide, new interrupt description, etc. but this doesn't explain all other modifications that do not change the functionality but rather naming, code location, etc. I can see the differences such as using

desc.cmd_desc_movi.id = girq->gi_irq - girq->gi_its_dev->lpis.lpi_base;

in movi command instead of i.e.

irq = irq - its_dev->lpis.lpi_base;
its_cmd_movi(sc, its_dev, col, irq);

but that is not something that cannot be applied on-top of the current ITS implementation, don't you think?

I will refrain from making further comments for separate lines or code blocks, I think what we all see here is clear.
Basically I don't understand this methodology to create a new file while adding support for INTRNG :-).
As far as I know (please correct me if I'm wrong), for INTRNG we need a set of callbacks such as enable interrupt, disable interrupt, post-filter, pre-ithread, etc. These are basically what we do now in GIC/ITS, if_pic.m so minor changes should be required to modify existing code and add missing routines required by INTRNG.
So how would you justify a need for a new file, moving the code around, changing names and definitions, etc.?

Very few files in the tree contain copyrights in the modules Makefiles. IMHO it would be better to just keep the minimum contents ($FreeBSD$ string) and that's all.
All the necessary copyrights are in the files that are build in the process.

Committed: https://svnweb.freebsd.org/changeset/base/299447

Removed unnecessary cache alignment

For "cpuset" command to make sense we should not bind CQ task to an arbitrary CPU.

I would add a comment there why we need two definitions of the same

Ups wrong version of the patch

I would fix comments before committing.
For example

/* Buffer must be larger than cache line for using cache zeroing
         * (and cache line aligned but this is checked after jump) */

to

/*
 * Buffer must be larger than cache line for using cache zeroing
 * (and cache line aligned but this is checked after jump)
 */

This patch works on VNIC

Wrong solution.