[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20260115132814.GF961588@nvidia.com>
Date: Thu, 15 Jan 2026 09:28:14 -0400
From: Jason Gunthorpe <jgg@...dia.com>
To: "Tian, Kevin" <kevin.tian@...el.com>
Cc: Baolu Lu <baolu.lu@...ux.intel.com>,
Samiullah Khawaja <skhawaja@...gle.com>,
Joerg Roedel <joro@...tes.org>, Will Deacon <will@...nel.org>,
Robin Murphy <robin.murphy@....com>,
Dmytro Maluka <dmaluka@...omium.org>,
"iommu@...ts.linux.dev" <iommu@...ts.linux.dev>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 3/3] iommu/vt-d: Rework hitless PASID entry replacement
On Thu, Jan 15, 2026 at 05:44:08AM +0000, Tian, Kevin wrote:
> or in Samiullah's case the old/new domains always contains the
> same mappings, so no corruption would ever occur, but that'd be
> a very KHO specific assumption. 😊
The thing witih KHO is we don't actually know this is true. We hope it
is true, but ultimately userspace is responsible to do it, and the
kernel doesn't check it.
This is why the hitless update in an appealing solution because it
means we don't need to deal with trying to adopt an entire page table
prepared by another kernel and ensuring this kernel has IOAS areas
that fully cover and exactly match what is in the table.
Unfortunately to do the extra DID will require a complex invalidation
sequence to hook multiple DIDs into the same domain
simultaneously.. Nicolin is working on this for ARM and I think
Intel's linked list scheme could be tasked to do it too..
Jason
Powered by blists - more mailing lists