[LU-19797] kfilnd use mempool for allocation inside kfilnd

• Type: Bug
• Resolution: Fixed
• Priority: Medium
• Fix Version/s: Lustre 2.18.0
• Severity: 3
• Assignee: Chris Horn
• Reporter: Chris Horn

Description

This change optimizes kfilnd transaction and message buffer
allocation to eliminate OOM-related sleeping, prevent filesystem
recursion deadlocks, and improve performance under memory
pressure.

Changed GFP_KERNEL to GFP_NOFS in critical allocation paths to prevent
filesystem recursion. The GFP_KERNEL flag can trigger memory reclaim
that calls into filesystem code, creating a deadlock when the
filesystem is waiting on network operations that are themselves trying
to allocate memory. The affected functions include
kfilnd_tn_alloc_common(), kfilnd_tn_set_sgl_buf(), kfilnd_peer_get(),
and kfilnd_ep_get_key().

Added mempool infrastructure to provide guaranteed allocation under
memory pressure. Two mempools are created: one for transaction objects
and one for immediate message buffers. Reserve sizes are calculated
dynamically based on peer_credits, number of CPTs, and a safety factor
of 2:
reserve = peer_credits * num_cpts * 2

The reserve sizes can be tuned via module parameters tn_reserve_min
and msg_reserve_min (default -1 enables auto-sizing). Mempool
statistics are exposed via debugfs
at /sys/kernel/debug/kfilnd/mempool_stats showing min_reserve,
curr_avail, and elements_allocated.

A kfilnd_tn_get_mempool_stats() accessor function was added to safely
expose mempool statistics to debugfs without breaking module
encapsulation.