Based on Alex's explanation [1] in the recent discussion let's update the comment explaining the memory lock limit calculation. [1] http://www.redhat.com/archives/libvir-list/2015-November/msg00329.html --- src/qemu/qemu_domain.c | 19 ++++++++++++++++--- 1 file changed, 16 insertions(+), 3 deletions(-) diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c index 12517a5..11cd39e 100644 --- a/src/qemu/qemu_domain.c +++ b/src/qemu/qemu_domain.c @@ -3639,9 +3639,22 @@ qemuDomainGetMlockLimitBytes(virDomainDefPtr def) goto done; } - /* VFIO requires all of the guest's memory to be locked resident, plus some - * amount for IO space. Alex Williamson suggested adding 1GiB for IO space - * just to be safe (some finer tuning might be nice, though). */ + /* For device pasthrough using VFIO the guest memory and MMIO memory regions + * need to be locked persistent in order to allow DMA. + * + * Currently the below limit is based on assumptions about the x86 platform. + * + * The chosen value of 1GiB below originates from x86 systems where it was + * the used as space reserved for the MMIO region for the whole system. + * + * On x86_64 systems the MMIO regions of the IOMMU mapped devices don't + * count towards the locked memory limit since the memory is owned by the + * device. Emulated devices though do count, but the regions are usually + * small. Although it's not guaranteed that the limit will be enough for all + * configurations it didn't pose a problem for now. + * + * Note that this may not be valid for all platforms. + */ memKB = virDomainDefGetMemoryActual(def) + 1024 * 1024; done: -- 2.6.2 -- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list