On 04/06/2017 08:37 AM, Leon Romanovsky wrote:
On Thu, Apr 06, 2017 at 07:45:16AM -0400, Dennis Dalessandro wrote:
On 04/06/2017 03:49 AM, Leon Romanovsky wrote:
On Wed, Apr 05, 2017 at 08:09:16PM +0000, Marciniszyn, Mike wrote:
> Is there a another way we should be looking at for setting things like this?
Use vendor channel interface to configure your driver.
What is that? configfs or something else?
An immediate answer without digging into your code is Matan's KABI work.
https://github.com/matanb10/linux/tree/abi-devel-latest
Until that code is formally accepted and actually in the kernel we can't
base our changes that are ready to go now (for 4.12) on it.
And we can't accept module parameters. I already presented my setup,
which I know in use by many people. Standalone kernel with everything
compiled in, everything runs in read-only small image without distro bloat.
it gives very small footprint, very fast execution and system
protection.
In such case, we'll be required to rebuild whole image to update command
line for one module parameter and we will need to do it just for
specific application, which is insane.
In the very rare case that if you care more about making mem dereg
faster at the expense of the data path you would have to do just that.
But for the vast majority of use cases the default is what you want,
keep performance benefits to the data path at the cost of memory dereg.
I'm glad that you realize now the importance of Matan's work and how it
can help overcome your current problems, You (Intel) are invited to help
him to make it faster.
This is the same stuff that a year ago was claimed it would only take a
couple weeks. In my opinion it's still more than one release out. When
it's done and Linus has accepted it, it's a different story.
However, I have an question, how do you ensure that user memory has no
users without refcounts? Will it be possible to dereg the memory despite
the fact that there are users?
Mike can correct me if I'm wrong but it is still refcounted. Just not per
CPU, global if you will.
IMHO, global will be always more expensive than percpu, due to locality.
However your patch presents different picture. You are claiming that removing
percpu_refcnt and leaving global will make work faster. How will it be?
Mike can explain better I'm sure but the gist of it is there is an
implicit RCU delay waiting for the async per CPU model to quiesce ref
counts to zero in the de-reg.
So while in the per CPU case, other things can be going on which helps
packet reception and posting of sends, the benefit to the data path.
However this comes at a cost to the de-reg.
So in the rare event that you want the de-reg to be faster and are
willing to take the data path hit, it's better to do it atomically
across all CPUs.
-Denny
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
