Skip to main content

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
Re: [paho-dev] Design question - how to best approach background tasks for the embedded clients

> FYI, Richard replied with the following:

I though my replies were going to the list rather than you directly! If not, I would appreciate it if you could also forward my reply to your second email.

Regards,
Richard.

+ http://www.FreeRTOS.org
Designed for simplicity. More than 113000 downloads in 2014.

+ http://www.FreeRTOS.org/plus
IoT, Trace, Certification, FAT FS, TCP/IP, Training, and more...



On 29/05/2015 10:54, Ian Craggs wrote:
FYI, Richard replied with the following:

"Are the two tasks at the same priority?  If not, then the higher
priority of the two will always get the semaphore.  If this is the
scenario you have, and you need the other task to get the semaphore too,
then let me know.

If on the other hand the two tasks are running at the same priority, and
the tasks take the mutex almost straight after giving it, then you can
find that the opportunity for a switch between giving and taking is
small enough that it appears that the same task always gets it.  If this
is the scenario you have, then call taskYIELD() after giving the mutex.
  That will force a switch to the other task and should work for you.
Older versions of FreeRTOS used to do that automatically, but it could
lead to thrashing switches, so the behaviour was changed."

I changed the priority of the background task to be less than or equal
to the task starting it (the application task in this case) and
everything started working as expected.

Ian


On 05/28/2015 09:32 PM, Ian Craggs wrote:
Hi Richard,

in implementing synchronization between two tasks in the FreeRTOS
Windows "simulator" I'm having a problem.  I'm starting a background
MQTT task -- that works fine.  Then I'm trying to use a mutex for
mutual exclusion between the two tasks.

I create the mutex with xSemaphoreCreateMutex().

One task successfully locks the mutex with SemaphoreTake().

The second task attempts to lock the mutex with SemaphoreTake() and
blocks.

The first task then calls SemaphoreGive() to release the mutex.

But the second task is not unblocked before the first task locks the
mutex again with SemaphoreTake().

Do I need to do something to make sure a task switch occurs when
SemaphoreGive() is called?

Thanks

Ian

On 05/25/2015 05:50 PM, Richard Barry wrote:
Hi Ian,

For ease of use and maintainability reasons it would, in my humble
opinion, be best to not only hide the cycle() function, but also the
yield() function.  That would mean, if yield were called at all, it
would be inside the public API functions.

I am still not completely clear as to the purpose of yield(), but
suspect if a separate MQTT thread were allowed then yield() would not
be required.  A background thread would free up processing time by
being event driven (important in low power systems that want to
sleep), so improve run time resource usage.  If it is considered that
there is not enough RAM resource for the background thread then
consider how much RAM could be saved in each thread using MQTT if the
RAM were instead spent in a single MQTT thread.  If it were still a
problem then there is already a TCP/IP thread, so maybe that could be
enhanced to allow user defined extensions - then run the MQTT stuff
in there [that last suggestion would not be a quick change!].

I would envisage an application thread using the MQTT thread as a
sort of proxy server.  The application thread just sends data via the
MQTT thread - all the ins and outs of how the data is sent, and QoS
level maintained, would be entirely inside the MQTT thread itself.
The application thread can sleep until notified by the MQTT thread as
to the the success or failure of the send. Likewise when receiving
data, the receive can be handled inside the MQTT thread, and then a
callback function (or queue, or whatever) used to send received data
to application threads that need it.

If you can we suggest one send or receive scenario to concentrate on,
and describe the sequence used to implement the scenario that
involves calls to yield(), then perhaps I could follow up with a
suggestion of how it could possibly be implemented without the yield().

Regards,
Richard.

+ http://www.FreeRTOS.org
Designed for simplicity. More than 113000 downloads in 2014.

+ http://www.FreeRTOS.org/plus
IoT, Trace, Certification, FAT FS, TCP/IP, Training, and more...



On 22/05/2015 12:36, Ian Craggs wrote:
Hello all,

as I've been updating the C "high level" embedded client for FreeRTOS:

(http://git.eclipse.org/c/paho/org.eclipse.paho.mqtt.embedded-c.git/commit/?h=develop)



I've been pondering how to best have background tasks completed.   I
originally envisaged two styles of APIs, similar to the standard Paho C
client libraries, synchronous and asynchronous.  I haven't yet created
the asynchronous version, but that would use a background thread, at
the
cost of using more resources.

The "synchronous" or blocking API would be easy to use, and all
processing would take place on the application's thread. Because no
separate threads were involved, this would both simplify implementation
and reduce resource usage.  To allow messages to be received, there
is a
yield() call, which takes as its parameter a time to run.  The API also
has a a function (or method) called cycle(), which processes one
incoming MQTT packet.  Yield() just calls cycle() under the covers.  I
didn't originally intend to make cycle() a public function.

The problem with yield() is that it might spend most of its time doing
nothing.  An MQTT packet might arrive at the end of the timeout period,
and not get processed until the next call to yield().   One solution is
to make cycle() a public function and allow the application to
interleave application with MQTT processing at a fine-grained level.

For FreeRTOS I was considering also a background task to be started.
But
this would still have the negative effects of increased resource usage
and complexity as synchronization of the common data structures
would be
needed.  So I'm reconsidering this once more, and thinking of leaving
the background thread to an (almost) totally non-blocking API.

I had also been thinking of changing the semantics of yield() in some
way to be more event driven.  But I don't know whether this would buy
any advantages over making cycle() a public function.

Thoughts?

_______________________________________________
paho-dev mailing list
paho-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or
unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/paho-dev




Back to the top