Chapter 36. Kernel

Some ext4 file systems cannot be resized

Due to a bug in the ext4 code, it is currently impossible to resize ext4 file systems that have 1 kilobyte block size and are smaller than 32 megabytes.

Repeated connection loss with iSER-enabled iSCSI targets

When using the server as an iSER-enabled iSCSI target, connection losses occur repeatedly, the target can stop responding and the kernel becomes unresponsive. To work around this issue, minimize iSER connection losses or revert to non-iSER iSCSI mode.

Installer does not detect Fibre Channel over Ethernet disks on EDD systems

On EDD systems, FCoE disks are not detected automatically by Anaconda due to the edd driver missing. This makes such disks unusable during the installation.
To work around this problem, perform the following steps:
* Add fcoe=edd:nodcb to the kernel command line during the installation, the FCoE disks will be detected by anaconda.
* Add fcoe=edd:nodcb to the rescue image and boot the system with it.
* Add the edd module to the initrd image by executing the following commands:
#dracut --regenerate-all -f
#dracut --add-drivers edd /boot/initramfs-3.10.0-123.el7.x86_64.img
* Reboot the system with the default boot menu entry

NUMA balancing does not work optimally under certain circumstances

The Linux kernel Non-Uniform Memory Access (NUMA) balancing does not work optimally under the following condition in Red Hat Enterprise Linux 7. When the numa_balancing option is set, some of the memory can move to an arbitrary non-destination node before moving to the constrained nodes, and the memory on the destination node also decreases under certain circumstances. There is currently no known workaround available.

PSM2 MTL disabled to avoid conflicts between PSM and PSM2 APIs

The new libpsm2 package provides the PSM2 API for use with Intel Omni-Path devices, which overlaps with the Performance Scaled Messaging (PSM) API installed by the infinipath-psm package for use with Truescale devices. The API overlap results in undefined behavior when a process links to libraries provided by both packages. This problem affects Open MPI if the set of its enabled MCA modules includes the psm2 Matching Transport Layer (MTL) and one or more modules that directly or indirectly depend on the library from the infinipath-psm package.
To avoid the PSM and PSM2 API conflict, Open MPI's psm2 MTL has been disabled by default in the /etc/openmpi-*/openmpi-mca-params.conf configuration file. If you enable it, you need to disable the psm and ofi MTLs and the usnic Byte Transfer Layer (BTL) that conflict with it (instructions are also provided in comments in the configuration file).
There is also a packaging conflict between the libpsm2-compat-devel and infinipath-psm-devel packages because they both contain PSM header files. Therefore, the two packages cannot be installed at the same time. To install one, uninstall the other.

Performance problem of the perf utility

The perf archive command, which creates archives with object files with build IDs found in files, takes a long time to complete on IBM System z. At present, no known workaround exists. Other architectures are not affected.

qlcnic fails to enslaved by bonding

Certain bonding modes set a MAC address on the device which the qlcnic driver does not properly recognize. This prevents the device from restoring its original MAC address when it is removed from the bond.
As a workaround, unenslave the qlcnic driver and reboot your operating system.

Installation fails on some 64-bit ARM Applied Micro computers

Red Hat Enterprise Linux 7.2 fails to install on certain 64-bit ARM systems by Applied Micro with the following error message:
Unable to handle kernel NULL pointer dereference at virtual address 0000033f
At present, there is no workaround for this problem.

libvirt management of VFIO devices can lead to host crashes

The libvirt management of host PCI devices, assigned to guests using the VFIO driver, can lead to host kernel drivers and the vfio-pci driver binding simultaneously to devices in the same IOMMU group. This is an invalid state, which can lead to a host unexpected termination.
For now, the only workaround is to never hot-unplug a VFIO device from a guest, if there are any other devices in the same IOMMU group.

Installation using iSCSI and IPv6 hangs for 15 minutes

Dracut times out after trying to connect to the specified iSCSI server for 15 minutes if IPv6 is enabled. Eventually, Dracut connects successfully and proceeds as expected; however, to avoid the delay, use ip=eth0:auto6 on the installer's command line.

i40e NIC freeze

With old firmware, a network card using the i40e driver becomes unusable for about ten seconds after it enters the promiscuous mode. To avoid this problem, update the firmware.

i40e is issuing WARN_ON

The i40e driver is issuing the WARN_ON macro during ring size changes because the code is cloning the rx_ring struct but not zeroing out the pointers before allocating new memory. Note that this warning is harmless to your system.

netprio_cgroups not mounted at boot

Currently, systemd mounts the /sys/fs/cgroup/ directory as read-only, which prevents the default mount of the /sys/fs/cgroup/net_prio/ directory. As a consequence, the netprio_cgroups module is not mounted at boot. To work around this problem, use the mount -o remount command, followed by rw -t cgroup nodev /sys/fs/cgroups. This makes it possible to install module-based cgroups manually.