core: name both kernel changes behind the EPOLLEXCLUSIVE gate

The comments and docs said the woken-but-not-running problem that got
EPOLLEXCLUSIVE disabled in 3596efd was fixed in Linux 6.0. The main
fix is 412895f03cbf in 5.7, which removes a woken waiter from the epoll
wait queue when the wake-up succeeds; 6.0 (a16ceb139610,
ep_autoremove_wake_function) removes it even when the wake-up fails.
Say so, and that the 6.0 gate stays because 5.7-5.19 were not tested.
Comments and documentation only.

Author: Lt-Flash

docs/manual/Script-CoreParameters.md

@@ -1275,7 +1275,7 @@ How TCP main hands a received message to the TCP workers. The default value is `
 
 * `shared` - all TCP workers wait on one shared dispatch socket. Every dispatched message wakes every idle worker (one of them wins the read, the others return empty-handed), and the kernel wakes the last-forked worker first, so the per-worker load forms a descending ladder from the last worker to the first. This was the only behaviour before this parameter existed.
 * `per_worker` - one dispatch socket per TCP worker; TCP main hands each message to the next idle active worker in round-robin order (or to the next active one when all are busy). One wakeup per message, even load across the workers.
-* `exclusive` - the shared socket, registered with `EPOLLEXCLUSIVE` so that only one waiting worker is woken per message. Allowed with the `epoll` reactor on Linux 6.0 and newer only; on older kernels a woken but not yet running worker keeps absorbing the following messages, so bursts serialize on it - OpenSIPS falls back to `shared` with a warning there.
+* `exclusive` - the shared socket, registered with `EPOLLEXCLUSIVE` so that only one waiting worker is woken per message. Allowed with the `epoll` reactor on Linux 6.0 and newer only, and OpenSIPS falls back to `shared` with a warning elsewhere. Before Linux 5.7 a woken but not yet running worker keeps absorbing the following messages, so bursts serialize on it; 5.7 removes a woken worker from the epoll wait queue at wake-up and 6.0 completes that, and kernels between the two were not tested.
 * `auto` - `per_worker`, on every platform (it only needs `AF_UNIX` datagram socket pairs, which the `shared` mode already uses).
 
 Example of usage:

globals.h

@@ -56,7 +56,8 @@ extern int udp_workers_no;
  *  shared    - one socket, every worker polls it (each packet wakes all idle
  *              workers, the last-forked one is woken first)
  *  exclusive - same socket, EPOLLEXCLUSIVE (one wakeup per packet; needs a
- *              6.0+ kernel, otherwise bursts serialize on one worker)
+ *              6.0+ kernel: before 5.7 bursts serialize on one worker,
+ *              5.7-5.19 are untested, see io_watch_add())
  *  reuseport - one SO_REUSEPORT socket per worker, a cBPF program spreads
  *              the packets uniformly at random (needs a 4.5+ kernel)
  *  auto      - reuseport when the kernel supports it, else shared */

io_wait.h

@@ -550,11 +550,15 @@ inline static int io_watch_add(	io_wait_h* h, // lgtm [cpp/use-of-goto]
 again1:
 #ifdef EPOLLEXCLUSIVE
 				/* opt-in only (IO_WATCH_EXCLUSIVE); the unconditional use
-				 * was disabled in 3596efdc00 because, with level-triggered
-				 * epoll on kernels before 6.0, a woken but not yet running
-				 * process keeps absorbing the following events (fixed by
-				 * ep_autoremove_wake_function) - callers must gate on the
-				 * kernel version, see os_kernel_at_least() */
+				 * was disabled in 3596efdc00 because a woken but not yet
+				 * running process kept absorbing the following events: it
+				 * stayed on its epoll's wait queue until it ran, so
+				 * ep_poll_callback() still saw a waiter and ended the
+				 * exclusive wake-up there. Linux 5.7 (412895f03cbf) removes
+				 * the waiter when the wake-up succeeds, 6.0 (a16ceb139610,
+				 * ep_autoremove_wake_function) even when it fails. 5.4
+				 * serializes bursts, 6.12 does not, 5.7-5.19 are untested,
+				 * so callers gate on 6.0, see os_kernel_at_least() */
 				if (excl && (flags & IO_WATCH_READ))
 					ep_event.events|=EPOLLEXCLUSIVE;
 #endif