Ensure RDMA service loads modules in initrd #1481
Conversation
Ensure that the rdma-load-modules@.service is started as part of the initrd. Fixes: 39fa824 ("redhat: add udev/systemd/etc infrastructure bits") Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
Ensure that the rdma-load-modules@.service is started as part of the initrd. Fixes: 7752410 ("suse: fix dracut support") Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
rdma-load-modules@.service run in the initrd. However, it gets terminated when initrd-cleanup.service isolates for initrd-switch-root.target. The termination can occur in the middle of the IPoIB initialization, leading to a failure to load netdevices. Include 'Before=initrd.target' to ensure that the services are not being killed when initrd-cleanup.service isolates to initrd-switch-root.target. Kernel log: workqueue: Failed to create a rescuer kthread for wq "ipoib_wq": -EINTR Cleaning Up and Shutting Down Daemons ib0: failed to allocate device WQ mlx5_0: failed to initialize device: ib0 port 1 (ret = -12) mlx5_0: couldn't register ipoib port 1; error -12 workqueue: Failed to create a rescuer kthread for wq "mlx5e": -EINTR ibp6s0f1, 1: ipoib_intf_alloc failed -12 workqueue: Failed to create a rescuer kthread for wq "mlx5e": -EINTR ibp6s0f2, 1: ipoib_intf_alloc failed -12 workqueue: Failed to create a rescuer kthread for wq "mlx5e": -EINTR ibp6s0f3, 1: ipoib_intf_alloc failed -12 Stopped Load RDMA modules …/rdma/modules/infiniband.conf Stopped Load RDMA modules …m /etc/rdma/modules/rdma.conf Fixes: 2f4fb9f ("Common infrastructure for auto loading rdma modules") Signed-off-by: Carolina Jubran <cjubran@nvidia.com>
|
Can i ask why are you running this service in initrd? |
We run this service in initrd to load RDMA modules early in the boot process. It prevents issues like failed NFS mounts over RDMA due to missing modules. |
|
I talked to Chuck and he told me that NFS module autoloading works fine, even checked it out for me. So why is it failing in the initrd? |
The rdma service is terminated when initrd-cleanup.service isolates for initrd-switch-root.target because initrd.target lacks the dependency on the rdma service. |
|
Does this actually work out of the box? I fill like SUSE/RH spec file (at least) are missing a call to dracut somewhere? |
I don't think so, we saw this failure in our regressions where we run dracut anyway. |
Fixed an issue where the RDMA service was killed during switch to root. Added a wants symlink for initrd.target in the dracut arrangement and Before=initrd.target to the systemd service to ensure it runs and completes during initrd.