On Fri, Mar 04, 2022 at 12:36:17PM -0500, Boris Ostrovsky wrote:

quoted

quoted

I bisected it to "x86: remove the IOMMU table infrastructure" but haven't actually looked at the code yet.

That looks like the swiotlb buffer did not get initialized at all, but I
can't really explain why.

Can you stick in a printk and see if xen_swiotlb_init_early gets called
at all?

Actually, that's the only thing I did do so far and yes, it does get called.

So, specifically for "x86: remove the IOMMU table infrastructure" I
think we need the one-liner below so that swiotlb_exit doesn't get called
for the Xen case.  But that should have been fixed up by the next
patch already.

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 2ac0ef9c2fb76..1173aa282ab27 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c

@@ -70,7 +70,7 @@ static void __init pci_xen_swiotlb_init(void)
 	if (!xen_initial_domain() && !x86_swiotlb_enable &&
 	    swiotlb_force != SWIOTLB_FORCE)
 		return;
-	x86_swiotlb_enable = false;
+	x86_swiotlb_enable = true;
 	xen_swiotlb = true;
 	xen_swiotlb_init_early();
 	dma_ops = &xen_swiotlb_dma_ops;