sockets_mgm: fix race in receive_fd causing infinite loop on reload

[NormB]

Summary

Fixes #3789 — sockets_mgm reload causes all OpenSIPS worker processes to spin at 100% CPU in an infinite loop inside push_sock2list() → sock_listadd().

Root Cause

During sockets_reload, every process receives an IPC RPC to run rpc_socket_reload_proc(). Worker (non-dynamic) processes close their copy of each dynamic socket, then call receive_fd() on the shared sock_mgm_unix[0] socketpair to obtain a fresh fd from the mgm process.

Since sock_mgm_unix[0] is a single SOCK_STREAM socket shared across all worker processes, concurrent receive_fd() calls race: worker A can consume the fd+metadata response intended for worker B. When this happens, worker B receives the wrong struct sock_mgm * pointer — one that references a socket already present in worker B's listener list. The sock_listadd() macro then corrupts the linked list into a circular loop (si->next == si), causing push_sock2list() to spin indefinitely.

GDB confirmed: si->next pointed to itself, and protos[1].listeners formed an infinite cycle through the same node.

Fix

Add a sock_mgm_reload_lock that serializes the entire send-IPC-to-mgm + receive-fd sequence across worker processes, ensuring only one worker at a time performs the fd-passing handshake on the shared socketpair.

Dynamic (mgm) processes are excluded from this lock because they create sockets directly via sock_mgm_add_listener() and never call receive_fd(). Including them would deadlock: the worker holding the lock blocks on receive_fd() waiting for the mgm process to handle rpc_sockets_send(), but the mgm process would be spinning on the same lock.

Changes

• sockets_mgm.c: Add sock_mgm_reload_lock (shared memory lock), allocated and initialized in mod_init(). In rpc_socket_reload_proc(), non-dynamic processes acquire the lock before the reload sequence and release it after all receive_fd() calls complete.

Test plan

• Start OpenSIPS with 2 dynamic UDP sockets from DB — 22 processes, 16 dynamic socket entries, all sleeping
• 1st sockets_reload (add 3rd socket) — MI returns OK, 24 socket entries, no CPU spinning
• 2nd sockets_reload (add 4th socket) — MI returns OK, 32 socket entries, no CPU spinning
• All processes remain in S (sleeping) state throughout — verified with ps and top
• Clean shutdown after testing