If fastReceiver() took more than 0.01 seconds to exit,
sleepySender() might have pushed a second message onto
the queue after setting recvExit, so there would be an
extra message in the queue for the next test, which I
was seeing on Appveyor. That's my current theory...
The internal mySend() and myReceive() routines do expect a timeout
of -1 to mean wait forever, see the epicsMessageQueueSend() and
epicsMessageQueueReceive() API routines.
Move the code that wakes up the next sending task to after we've
added our threadNode to the receiveQueue. He still has to wait for
us to release the Mutex though, so this might make no difference.
This commit also changes when we decrement the number of waiting
senders so it always happens immediately after a threadNode gets
taken off the sendQueue by the code that removed it.
When sending a message, if the queue is full so we have to wait, we
create a threadNode with an eventNode in it and stick it on the
sendQueue, then wait for a receiver to signal that event, waking us.
If we awoke due to a timeout but a receiver was actually waking us
up anyway (i.e. eventSent was set), we shouldn't give up.
Introduced freeEventNode() which ensures eventNodes don't have a
signalled event in them before returning the node to the freeList.
Callers pass the status from epicsEventWaitWithTimeout() to indicate
whether it was signalled or not. If it timed out we must trigger it
and Wait to clear the event state.
It appears that previously a negative timeout actually implemented a
'wait forever', but the VxWorks and RTEMS implementations both check
for (timeout <= 0) and return immediately if nothing can be done
without waiting.
Lists the directories with failed tests at the end of the build.
It is no longer necessary to use 'make -k' to see the results
of all tests after one or more failures as only the top-level
test-results recipe will generate a build error.
I split his two "Fields Common to ..." sections back into separate docs,
added links between them all, and made the appropriate build changes.
Also added these and the aai/aao records to the documentation index.
Also improve behavior in case signals are delivered to the sleeping
thread. This fixes a potential security weakness reported by codacy
(interaction of usleep with SIGALRM and other timer functions such
as sleep(), alarm(), setitimer(), and nanosleep() is unspecified).
This fix apply to event record device with constant INP.
Now when the event record is proccessed the associated records with the
same SCAN setup get triggered correctly, it is not more necessary to set
VAL on event record.
Fixes lp: #1829770
Working with Dirk Zimoch @dirk.zimoch, fixed various issues
with bit operations on VS2019 32bit. These seem to relate to
handling bit 31 of a 32 bit number.
As EPICS << is an arithmetic bit shift, following Java we
have added <<< and >>> operators for logical shifts
Though it is on a different architecture, this looks like
a similar issue to LP: #1838792
