Release 3.9.6

Major

duroslight: Added a sanity check on objects read from the primary to ensure that the rebuild read response data sizes comply with the replication protocol, even if the original object is malformed. This prevents link failures caused by sendmsg: Bad Address errors on Alma9 kernels.

LBM1-37887

Major

duroslight: Ensured that the primary adheres to protocol assumptions when passing objects, and that the secondary sanity checks received objects - detecting and correcting any violations so malformed objects can still be safely persisted. Previously, when a primary read an object requested by a secondary rebuild, it might have detected corruption (e.g., decompression failure, CRC mismatch). However, if dek_generation was corrupted (non-zero), the primary may have missed the issue and treated the object as encrypted, passing it as is. The secondary then assumed that the object was valid and writable, which could cause problems if the object was malformed.

LBM1-37650

Minor

duroslight: Fixed a bug that could cause slightly longer than needed rebuilds due to missed synchronization points, causing a rebuild based on an earlier synchronization point.

LBM1-35565

Minor

node-manager: Increased the default LimitNOFILE the node-manager sets in the duroslight-*.service files from 65536 to 300,000. This is to prevent duroslight process failure due to "Too many open files" with a large number of connected NVMe clients.

LBM1-36478

Minor

node-manager: Workaround Alma 9.4 bug that in some cases returned numa node -1 for NVMe devices.

LBM1-33157

Minor

pmem_init: Improved checking and (if necessary) inferring of the NUMA node.

LBM1-35890

Minor

userlbe: Fixed an issue where latency metrics were all bunched together in the smallest histogram bucket, making it harder to debug disk latency issues. Metrics are now correctly divided into buckets.

LBM1-35882