NOTE: Previously sent to the IOMMU list; see <http://thread.gmane.org/gmane.linux.kernel.iommu/3946/focus=3969>. That was in 3.12.

Linux 3.12 added a quirk to disable VT-d due to an Intel errata. However, having VT-d disabled on this machine (whether by the BIOS setting or by the kernel disabling it, all the permutations are in that email thread) results in IRQ errors like these:

[70090.499038] do_IRQ: 2.104 No irq handler for vector (irq -1)
[80893.648397] do_IRQ: 2.109 No irq handler for vector (irq -1)
[81293.703732] do_IRQ: 3.79 No irq handler for vector (irq -1)
[83814.393615] do_IRQ: 3.188 No irq handler for vector (irq -1)
[85504.918101] do_IRQ: 1.229 No irq handler for vector (irq -1)
(etc.)

Since 3.12, I've been building kernels with a local patch to leave VT-d enabled, which gets rid of the IRQ errors (and, also, the machine is perfectly stable).

Tests of 3.12 w/ both VT-d settings in bios and nointremap are at <http://news.gmane.org/find-root.php?message_id=52D83BF9.7040502%40metrics.net>; note the test dmesg's are in 1-3-2 order.