GitLab

With this rework, we have two separate per-thread mode flags for
greater flexibility aimed at controlling the debug features on a
per-thread basis:

- T_WOSS can be set to trigger SIGDEBUG upon (unexpected) stage switch
  to in-band mode. This is strictly equivalent to the obsoleted T_WARN
  bit.

- T_WOLI enables/disables the detection of locking inconsistencies
  with mutexes via the EVL_THRIOC_{SET, CLEAR}_MODE interface. This
  combines the former static CONFIG_EVL_DEBUG_MUTEX_INBAND and
  CONFIG_EVL_DEBUG_MUTEX_SLEEP options.

Enabling CONFIG_EVL_DEBUG_WOLI turns on T_WOLI by default for every
new EVL thread running in userland, which can be opted out on a
per-thread basis using EVL_THRIOC_CLEAR_MODE.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

The new hierachy of priority scales is as follows:

EVL_CORE_MIN_PRIO = EVL_WEAK_MIN_PRIO
...
	EVL_FIFO_MIN_PRIO ==
	     EVL_QUOTA_MIN_PRIO ==
	          EVL_TP_MIN_PRIO (== 1)
	...
	EVL_FIFO_MAX_PRIO =
	     EVL_QUOTA_MAX_PRIO ==
	          EVL_TP_MAX_PRIO ==
		       EVL_WEAK_MAX_PRIO (< MAX_USER_RT_PRIO)
...
EVL_CORE_MAX_PRIO (> MAX_RT_PRIO)

We reserve a couple of priority levels above the highest inband
kthread priority (MAX_RT_PRIO..MAX_RT_PRIO+1), which are guaranteed to
be higher than the highest inband user task priority
(MAX_USER_RT_PRIO-1) we use for SCHED_FIFO. Those extra levels can be
used for EVL kthreads which must top the priority of any userland
thread.

SCHED_EVL was dropped int the process, since userland is now
constrained to EVL_FIFO_MAX_PRIO by construction.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Requires a couple of ABI changes.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

We don't plan for switching CPU from out-of-band context anymore, so
T_MOVED is basically pointless now.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

This protocol operates the monitor value as a set of boolean event
flags forming a 32bit-wide event group. All bits from a group are
initially set to zero.

An event flag group is a lightweight notification mechanism. The
application can send bitmasks to raise individual bits from the group
(i.e. group_value |= bits), or wait for the group to have at least one
bit set for satisfying the request. In the latter case, the group
value is read then cleared atomically, and the collected bits are
returned to the thread heading the wait queue.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Introduce a 'protocol' specifier for monitors along with a basic type
(gate or event), which userland can pass to change the way a monitor
can be signaled and awaited. Define PI/PP gates using this new
information bit associated to the generic EVL_MONITOR_GATE type.

With this change in, the pre-existing ungated event can be enriched
with additional protocols beyond the semaphore-based P/V operations.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>

We may collect negative delta values when sampling latency on fast
machines, make sure the delta sum which is used in calculating the
average is signed, to prevent weird results.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Make sure breakage and/or WARN_ON() in the in-band transition code we
use for fixing up a trap on the oob stage cannot cause
handle_oob_trap() to be called recursively, leading to a stack
overflow eventually.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Jorge Ramirez-Ortiz authored Mar 28, 2019 and Philippe Gerum committed Jun 05, 2021

Elements:
--------
Basic features that cant be implemented from userland sharable between
processes.

Threads, monitors, xbufs and clocks -> every instance is backed by a
device in /dev/evl which allows sharing.

Utilities
---------
poll is a private-process construct; all poll groups are obtained from a
single multiplex device: /dev/evl/poll. They cannot be shared.

timers are in essence timerfd directly obtained from clocks, they don't
exist on their own, and are process-local too.

proxies are backed by a device in /dev/evl to allow for sharing, but
this is not something essential.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

The hectic utility needs to tell us whether some rtus/rtup/rtuo
threads are about to start waiting for an out-of-band event, so that
they don't miss a signal from the sleeper/switcher.

This information is now passed at UTASK registration time, which
ensures that HECTIC_OOB_WAIT is present in the task base flags early
enough before a SWITCH_TO request may target them.

Otherwise, the first evl_raise_flag() might be skipped in absence of
HECTIC_OOB_WAIT in the base flags of the target task, when the
sleeper/switcher manages to issue SWITCH_TO before such task enter the
PEND routine, which was seen at several occasions on very slow
platforms with 4+ CPUs (e.g. kvm-less qemu).

A typical test case involved running the following commands in
parallel on a 4-way configuration:

$ latmus
$ hectic rtuo2 rtuo3

In that case, tasks on CPU3 would miss the initial signal from the
sleeper/switcher.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>