Re: [PATCH] ahci: force 32-bit DMA for JMicron JMB582/JMB585

[Lennert Buytenhek <kernel@wantstofly.org>]

On Fri, Sep 04, 2026 at 11:55:38PM +0700, Mikael Etienne wrote:

> The EFR field is probably the most direct discriminant. If Arthur,
> Alvin and Lennert posted the same line, the "different IOMMU versions"
> hypothesis could be tested immediately, with no extra hardware:

[...]

To echo what Niklas said, the DMA problem I was seeing a while ago with
the ASM106x was purely an issue with the ARM106x itself, that is, a 
device-internal IOVA address width restriction, and had nothing to do
with the IOMMU in my system.

I was seeing AMD IOMMU I/O page faults in my particular case, but only
because the system I originally encountered the issue in was an AMD-based
system where the ASM106x happened to be behind an (AMD) IOMMU which was
by default enabled in an isolating fashion, and the device was indeed
causing I/O page faults, due to an IOVA address with restriction in the
device itself, and the AMD IOMMU correctly signaled those I/O page faults,
and the AMD IOMMU was not at fault for the device causing those I/O page
faults, and I did subsequently reproduce the issue on an Intel-based
system, where the ASM106x was causing the exact same I/O page faults --
that is, after the Intel IOMMU was manually enabled, which it wasn't
(and maybe still isn't?) by default.

In other words, I think that the issue I was seeing with the ASM106x
should probably be left out of consideration in the context of the current
discussion, where it appears from a distance that the (AMD) IOMMU in the
reporter's system is somehow implicated in the corruption and misbehavior
they are seeing.