[Xenomai-core] Re: RT-CAN tx_sem

[Wolfgang Grandegger <wg@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 2329 bytes --]

Jan Kiszka wrote:
> Wolfgang Grandegger wrote:
>> Hi Jan,
>>
>> Jan Kiszka wrote:
>>> Hi Wolfgang,
>>>
>>> fiddling with the CAN utils, I noticed the behaviour that rtcansend on
>>> freshly registered but stopped devices simply blocks. And when you
>>> meanwhile call rtcanconfig up on that device, rtcansend will continue to
>>> block.
>> I see the expected behavior on stopped devices:
>>
>>   bash-2.05b# rtcansend rtcan0
>>   rt_dev_send: Network is down
>>
>> This is, because the tx_sem is marked "destroyed" already when the
>> device gets initialized. I wonder why your app blocks.
> 
> Unlikely, because the sem is _not_ marked destroyed on startup:
> 
> http://www.rts.uni-hannover.de/xenomai/lxr/source/ksrc/drivers/can/rtcan_dev.c#181

You missed

     /* Set dummy state for following call */
     dev->state = CAN_STATE_ACTIVE;
     /* Enter reset mode */
     rtcan_sja_mode_stop(dev, NULL);

in the init part of rtcan_sja1000.c and rtcan_mscan.c, ...

> 
> In case it makes a subtle difference: I tested with xeno_can_virt, but I
> think I saw the same effect on real SJA1000 hw as well.

... which is missing in rtcan_virt.c. I already thought moving it to the 
common part. The attached patch fixes that.

>>> The reason is that during startup of CAN devices the tx-semaphore is
>>> re-initialised while it is already set up during device registration.
>>> Re-init on an already initialised rtdm_sem is, say, undefined.
>> That makes definitely trouble. A CAN_MODE_START should only start the
>> device if it's not active. The attached patch fixes this. Another
>> possibility would be to force a CAN_MODE_STOP in case the device is
>> already active (=restart).
>>
>>> So rtdm_sem should better only be initialised/destroyed by the drivers.
>>> Trying to send on a shut down CAN device could be catched differently,
>>> e.g. via the device state. This likely needs more thoughts, take it as a
>>> note that $something should be done.
>> With the fix above I do not see further problems with the existing
>> implementation using the "destroyed" state to mark the device unavailable.
> 
> The problem of double init remains. I don't think that the sem state
> should be used for encoding the CAN device state towards potential senders.

As I see it, Sebastian's trick saves code and synchronization.

Wolfgang.


[-- Attachment #2: xenomai-rtcan-txsem2.patch --]
[-- Type: text/x-patch, Size: 1169 bytes --]

+ diff -u xenomai/ksrc/drivers/can/rtcan_dev.c.TXSEM xenomai/ksrc/drivers/can/rtcan_dev.c
--- xenomai/ksrc/drivers/can/rtcan_dev.c.TXSEM	2007-02-15 11:21:43.000000000 +0100
+++ xenomai/ksrc/drivers/can/rtcan_dev.c	2007-02-15 14:52:20.000000000 +0100
@@ -176,9 +176,9 @@
 
     rtdm_lock_init(&dev->device_lock);
 
-    /* Init TX Semaphore, will be destroyed forthwith
-     * when setting stop mode */
+    /* Init and destroy TX semaphore to mark TX unavailable. */
     rtdm_sem_init(&dev->tx_sem, 0);
+    rtdm_sem_destroy(&dev->tx_sem);
 #ifdef RTCAN_USE_REFCOUNT
     atomic_set(&dev->refcount, 0);
 #endif
+ diff -u xenomai/ksrc/drivers/can/rtcan_raw_dev.c.TXSEM xenomai/ksrc/drivers/can/rtcan_raw_dev.c
--- xenomai/ksrc/drivers/can/rtcan_raw_dev.c.TXSEM	2007-02-15 11:21:43.000000000 +0100
+++ xenomai/ksrc/drivers/can/rtcan_raw_dev.c	2007-02-15 14:16:16.000000000 +0100
@@ -193,7 +193,8 @@
     switch (request) {
     case SIOCSCANMODE:
 	mode = (can_mode_t *)&ifr->ifr_ifru;
-	if (dev->do_set_mode)
+	if (dev->do_set_mode &&
+	    !(*mode == CAN_MODE_START && CAN_STATE_OPERATING(dev->state)))
 	    ret = dev->do_set_mode(dev, *mode, &lock_ctx);
 	break;