The routine qemuDomainGetMemLockLimitBytes() has a couple tests to determine what the maximum limit of locked memory should be. If I start a domain without any vfio stuff, /proc/$PID/limits says the limit is 64KiB. If I start a guest with a vfio-ccw hostdev, the limit is now $GUEST_MEMORY + 1GiB. It doesn't matter if I have one hostdev or a thousand; it's always 1GiB more than the configured amount of guest memory. If I start a guest without any vfio devices, and hotplug that same vfio-ccw hostdev, the limit remains at 64KiB and I start getting I/O errors in the guest. This makes sense, since some of the I/O chains are long enough to exceed that page limit, and the host starts throwing errors. There is already code that adjusts this limit during the hotplug of vfio-pci devices, so patch 1 refactors that code so that it can be re-used on the mdev hotplug path in patch 3. Patch 2, meanwhile, adds some cleanup that I think is missing in the vfio-pci path, based on my read of the mdev path. $ grep locked /proc/83543/limits Max locked memory 65536 65536 bytes $ virsh attach-device guest scratch-ca8b.xml Device attached successfully $ grep locked /proc/83543/limits Max locked memory 3221225472 3221225472 bytes Eric Farman (3): qemu: Refactor the max memlock routine qemu: Reset the maximum locked memory on hotplug fail qemu: Adjust max memlock on mdev hotplug src/qemu/qemu_domain.c | 30 ++++++++++++++++++++++++++++++ src/qemu/qemu_domain.h | 2 ++ src/qemu/qemu_hotplug.c | 22 ++++++++++++---------- 3 files changed, 44 insertions(+), 10 deletions(-) -- 2.17.1 -- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list
