On Wed, May 19, 2021 at 01:01:25PM -0700, Marc Orr wrote:

On Wed, May 19, 2021 at 10:25 AM Greg KH [off-list ref] wrote:

quoted

On Wed, May 19, 2021 at 10:18:38AM -0700, Marc Orr wrote:

quoted

On Wed, May 19, 2021 at 10:03 AM Greg KH [off-list ref] wrote:

quoted

On Wed, May 19, 2021 at 09:42:42AM -0700, Jianxiong Gao wrote:

quoted

On Wed, May 19, 2021 at 1:11 AM Greg KH [off-list ref] wrote:

quoted

I still fail to understand why you can not just use the 5.10.y kernel or
newer.  What is preventing you from doing this if you wish to use this
type of hardware?  This is not a "regression" in that the 5.4.y kernel
has never worked with this hardware before, it feels like a new feature.

NVMe + SWIOTLB is not a new feature. From my understanding it should
be supported by 5.4.y kernel correctly. Currently without the patch, any
NVMe device (along with some other devices that relies on offset to
work correctly), could be broken if the SWIOTLB is used on a 5.4.y kernel.

Then do not do that, as obviously it never worked without your fixes, so
this isn't a "regression".

NVMe + SWIOTLB works fine without this bug fix. By fine I mean that a
guest kernel using this configuration boots and runs stably for weeks
and months under general-purpose usage. The bug that this patch set
fixes was only encountered when a user tried to format an xfs
filesystem under a RHEL guest kernel.

quoted

And again, why can you not just use 5.10.y?

For our use case, this fixes the guest kernel, not the host kernel.
The guest distros that we support use 5.4 kernels. We do not control
the kernel that the distros deploy for usage as a guest OS on cloud.
We only control the host kernel.

And how are you going to get your guest kernels to update to these
patches?  What specific ones are you concerned about?

RHEL ignores stable kernel updates, so if you are worried about them,
please just work with that company directly.

We support COS as a guest [1], which does base their kernel on 5.4
LTS. If these patches were accepted into 5.4 LTS, they would
automatically get picked up by COS.

[1] https://cloud.google.com/container-optimized-os

Then go work with that group to add this "required" set of new features
for your cloud systems that require this as again, I fail to see how
this is a "regression" at all.

Maybe I should just go rip out these from 5.10.y as well as it feels
like a _very_ platform-specific issue that you all are having here.

thanks,

greg k-h