Blackfin doesn't switch tasks, resets dispatch_necessary before querying it

[Kolja Waschk]

Problem xperienced with bf537Stamp BSP on real hardware (a board with ADSP-BF537 and just SDRAM and flash attached, only PLL, SDRAM, and UART are initialized):

The ticker sample starts its three tasks, but all stall forever in rtems_task_wake_after(x) for any x greater than 1. Only the Idle task runs.

The clockISR and rtems_clock_tick are executed and clock count is incremented.

Debugging via JTAG with breakpoint at _Thread_Delay_ends made me discover that:

1. _Thread_Dispatch_necessary = true is set by _Scheduler_priority_unblock

2. Then it is cleared in cpukit/score/cpu/bfin/cpu_asm.S:480 (doDispatch) by writing 0 to DISPATCH_NEEDED

3. Finally _Thread_Dispatch, called by _ISR15_Handler, doesn't do any Context_Switch because per_cpu->dispatch_necessary (the very same byte) is zero.

As a workaround, removing the two lines after doDispatch in cpu_asm.S leads to a running ticker sample.

doDispatch:

/* r0 = 0: */
/* B[p0] = r0; */
raise 15;

I have not tried other tests yet. I suspect some more inconsistencies. This doesn't appear like it ever worked...?

Also see discussion here:
http://www.rtems.org/pipermail/rtems-users/2013-October/012038.html

Another occurence with similar outcome, maybe for the same reason:
http://www.rtems.org/pipermail/rtems-users/2013-May/011544.html

Probably in RTEMS since 2008, with background e.g. discussed here:
http://www.rtems.org/pipermail/rtems-users/2008-April/003129.html

[Sebastian Huber]

The general interrupt with thread dispatch sequence is depicted here:

http://www.rtems.org/wiki/index.php?title=SMP#Thread_Restart

[Joel Sherrill]

I have no recollection of a patch being submitted. Does anyone else?

Also the URLs for mail point to the old mhonarc archives so I am having trouble finding what they are talking about.

[Kolja Waschk]

There is no patch yet. We produce devices that are running very stable now, with the modification as noted in the issue description (commenting out the two lines in cpu_asm.S). But I have not been able to correctly identify all possible side effects nor run all tests (due to other problems).

Updates for the links:

Also see discussion here ("ticker hangs on bf537Stamp"):
​http://lists.rtems.org/pipermail/users/2013-October/027186.html

Another occurence with similar outcome, maybe for the same reason ("bfin simulator problems"):
​http://lists.rtems.org/pipermail/users/2013-May/026692.html

Probably in RTEMS since 2008, with background e.g. discussed here ("Problems with the blackfin port"):
​http://lists.rtems.org/pipermail/users/2008-April/018266.html

[Gedare Bloom]

Replace Version=HEAD with Version=4.11 for the tickets with Milestone >= 4.11

[Gedare Bloom]

Bump priority to highest for tickets with a fix attached or seemingly simple fix proposed in the description or comments.

[Gedare Bloom <gedare@…>]

In [changeset:"7f1bb40aa1b87bc346bd07b893ebdd5c35e10410/rtems"]:

bfin: do not reset dispatch needed

Closes #2148

Fix suggested in above ticket. On examination, the assembly
appears to be clearing the DISPATCH_NEEDED flag before jumping
to _Thread_Dispatch.