Jan Kiszka wrote:
> Hi Philippe,
> Jan Kiszka wrote:
>>Just ran into this with CONFIG_IPIPE_DEBUG_CONTEXT (maybe due to some
>>bug of my own):
> Here is some code to trigger the issue reliably:
> #include <sys/mman.h>
> #include <native/task.h>
> void task_fnct(void *arg)
> RT_TASK task;
> rt_task_spawn(&task, "task", 0, 10, 0, task_fnct, NULL);
>>[ 102.616000] I-pipe: Detected illicit call from domain 'Xenomai'
>>[ 102.616000] into a service reserved for domain 'Linux' and below.
>>[ 102.616000] c741bdc8 00000000 00000000 c8860ef8 c741bdec c0105683
>>[ 102.616000] c0361f00 c741be08 c01519ed c032f5b8 c032c742 c03380b3
>>[ 102.616000] c741be14 c0142ce9 c7a80b30 c741be3c c884d075 c885f150
>>[ 102.616000] Call Trace:
>>[ 102.616000] [<c0104d9f>] show_trace_log_lvl+0x1f/0x40
>>[ 102.616000] [<c0104e71>] show_stack_log_lvl+0xb1/0xe0
>>[ 102.616000] [<c0105683>] show_stack+0x33/0x40
>>[ 102.616000] [<c01519ed>] ipipe_check_context+0xad/0xc0
>>[ 102.616000] [<c0142ce9>] module_put+0x19/0x90
>>[ 102.616000] [<c884d075>] xnshadow_unmap+0xb5/0x130 [xeno_nucleus]
>>[ 102.616000] [<c8871dc5>] __shadow_delete_hook+0x25/0x30 [xeno_native]
>>[ 102.616000] [<c8842f78>] xnpod_schedule+0xb58/0x12f0 [xeno_nucleus]
>>[ 102.616000] [<c8844bfb>] xnpod_delete_thread+0x2cb/0x3d0 [xeno_nucleus]
>>[ 102.616000] [<c886f5bd>] rt_task_delete+0x20d/0x220 [xeno_native]
>>I would dare to say that module_put in xnshadow_unmap is not well placed
>>as it can wakeup a Linux process. The module ref-counter maintenance
>>needs some postponing, I guess.
> Attached is a patch proposal. It solves the issue by postponing the
> module_put via a new schedule_linux_call. Note that this approach issues
> LO_WAKEUP_REQ where the old test (p->state != TASK_RUNNING) would not
> have done so. I don't see negative side effects yet, and I'm furthermore
> not sure of the old code was handling SMP scenarios safely (What if the
> thread to be unmapped was running on different CPU than xnshadow_unmap?
> How to ensure test-atomicity then?).
This one counts as mine! I am Ok with the fix, but IMHO, the
"if(p->state != TASK_RUNNING)" probably has a reason, so I would leave
it in the new implementation.
Xenomai-core mailing list