On 9/4/2025 4:49 PM, Robin Jarry wrote:
Hey Anatoly,
Burakov, Anatoly, Sep 04, 2025 at 16:35:
I have been trying to reproduce this but I'm unable to. I run some old
CentOS so I built GCC15 manually and tried reproducing with that. The
version I'm using is 15.2, could you please retest with latest GCC15 and
see if it was fixed?
I am running on Fedora 42 that has GCC 15.2.1 and the problem persists.
(if not, I'll try to get a more recent distro and reproduce again)
The DMA mask checking you're referring to is there for cases where e.g.
in a VM the emulated IOMMU will not have full address width (39 bits was
common at the time), and thus we couldn't use the full VA space for IOVA
addressing, and had to resort to using real physical addresses (because
kernel ensure those are to be within support range of IOMMU). Were you
running this test in a VM, by any chance?
No I am running on a physical machine. I tested again without --no-huge
and 2M/1G huge pages, and the issue persists as well.
I tried both as root, and as a regular user.
It should work if you use --iova-mode=pa (even for VM case), this will
require root though. However, I don't quite get how you get the error
that you do.
Specifically, it's weird that your DMA width is 42 bits (indicated by
"using dma mask fffffc0000000000" log), running on a physical machine -
mine sits at 57 bits, so despite having VA's within the same range as
yours, this works on my machine because my DMA mask is much wider. Maybe
there's a kernel command-line of some sort that limits your DMA width?
At first glance I don't think the issue has anything to do with DPDK, at
least it doesn't look like it does to me. I do not see how DMA mask
width would be affected by libasan or GCC15, to me it sounds more like a
kernel configuration/HW setup problem.
--
Thanks,
Anatoly
