check for IOMMU support on linux

I'd like to verify on any given Linux machine if PCI passthrough is supported. After a bit of googling, I found that I should rather check if IOMMU is supported, and I did so by running:

dmesg | grep IOMMU

If it supports IOMMU (and not IOMMUv2), I would get:

IOMMU                                                          
[    0.000000] DMAR: IOMMU enabled
[    0.049734] DMAR-IR: IOAPIC id 8 under DRHD base  0xfbffc000 IOMMU 0
[    0.049735] DMAR-IR: IOAPIC id 9 under DRHD base  0xfbffc000 IOMMU 0
[    1.286567] AMD IOMMUv2 driver by Joerg Roedel <[email protected]>
[    1.286568] AMD IOMMUv2 functionality not available on this system

...where DMAR: IOMMU enabled is what I'm looking for.

Now, if the machine has been running for days without a reboot, that first message [ 0.000000] DMAR: IOMMU enabled might not appear any more in the log with the previous command.

Is there any way to check for IOMMU support when that message disappears from the log?

Solution

Since 2014 enabled iommu are registered in /sys (sysfs) special file system as class iommu (documented at ABI/testing/sysfs-class-iommu): https://patchwork.kernel.org/patch/4345491/ "[2/3] iommu/intel: Make use of IOMMU sysfs support" - June 12, 2014

Register our DRHD IOMMUs, cross link devices, and provide a base set of attributes for the IOMMU. ... On a typical desktop system, this provides the following (pruned):
$ find /sys | grep dmar
/sys/devices/virtual/iommu/dmar0
...
/sys/class/iommu/dmar0
/sys/class/iommu/dmar1

The code is iommu_device_create (http://elixir.free-electrons.com/linux/v4.5/ident/iommu_device_create, around 4.5) or iommu_device_sysfs_add (http://elixir.free-electrons.com/linux/v4.11/ident/iommu_device_sysfs_add) in more recent kernels.

/*
 * Create an IOMMU device and return a pointer to it.  IOMMU specific
 * attributes can be provided as an attribute group, allowing a unique
 * namespace per IOMMU type.
 */
struct device *iommu_device_create(struct device *parent, void *drvdata,
                   const struct attribute_group **groups,
                   const char *fmt, ...)

Registration is done only for enabled IOMMU. DMAR:

if (intel_iommu_enabled) {
    iommu->iommu_dev = iommu_device_create(NULL, iommu,
                           intel_iommu_groups,
                           "%s", iommu->name);

AMD IOMMU:

static int iommu_init_pci(struct amd_iommu *iommu)
{ ...
    if (!iommu->dev)
        return -ENODEV;
...
    iommu->iommu_dev = iommu_device_create(&iommu->dev->dev, iommu,
                           amd_iommu_groups, "ivhd%d",
                           iommu->index);

Intel:

int __init intel_iommu_init(void)
{ ...
    pr_info("Intel(R) Virtualization Technology for Directed I/O\n");
...
    for_each_active_iommu(iommu, drhd)
        iommu->iommu_dev = iommu_device_create(NULL, iommu,
                               intel_iommu_groups,
                               "%s", iommu->name);

With 4.11 linux kernel version iommu_device_sysfs_add is referenced in many IOMMU drivers, so checking /sys/class/iommu is better (more universal) way to programmatically detect enabled IOMMU than parsing dmesg output or searching in /var/log/kern.log or /var/log/messages for driver-specific enable messages:

Referenced in 10 files:

drivers/iommu/amd_iommu_init.c, line 1640

drivers/iommu/arm-smmu-v3.c, line 2709

drivers/iommu/arm-smmu.c, line 2163

drivers/iommu/dmar.c, line 1083

drivers/iommu/exynos-iommu.c, line 623

drivers/iommu/intel-iommu.c, line 4878

drivers/iommu/iommu-sysfs.c, line 57

drivers/iommu/msm_iommu.c, line 797

drivers/iommu/mtk_iommu.c, line 581