Experimental Physics and
Industrial Control System

Peregrine McGehee <[email protected]> · Tue, 5 Dec 1995 16:31:16 -1000

We have been successfully running EPICS R3.12.1 on Sun Sparc 1e and
Force Sparc 5 based IOCs for the last few months and last week
encountered the following sequencer behavior:

At a period of several seconds to several minutes the Sequencer
Auxillary Task crashes with a memory alignment error. Also the vxWorks
routine memShow(1) forces a crash as well. 

The root cause of this was found to be:

If the sequencer application accessed any non-local process variables 
(i.e. actually used Channel Access) then the stack size on "seqAux"
would overflow the 2000 byte size defined in seq_main.c.

I suspect that this is due to the message buffers being padded out to
comply with 8-byte alignment. 

I haven't identified the specific section of code within channel access
that seems to be responsible for this growth of the stack but the
difference between using CA and using only DB access shows up in the CA
routine cac_select_io(). [technical note: iiuList.node.next is always
NULL if no CA (UDP or TCP) related file descriptors need to be checked].

So - the stats on the seqAux stack for the sparc port are:
CUR: 736	HIGH: 2848
The high-water-mark grows during the first few minutes and then seems
to be stable.

Note: the _seqAuxTask runs at one priority level above the actual
sequencer app - and it's _only_ function is to call
ca_pend_event(10.0) forever. Why is that?

Aloha,
	Peregrine

-------------------------------------------------------------------------------
Peregrine M. McGehee			Canada-France-Hawaii Telescope
Telescope Control Systems Group		P.O. Box 1597, Kamuela, Hawaii

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System