Re: stress-ng process termination issue

[Michael Kelly <mike@weatherwax.co.uk>]

On 31/07/2025 13:29, Michael Kelly wrote:
> I think that is possible and worth a try. The process termination is 
> currently slowed significantly after the high pageout during the test 
> run which is why the SIGKILL is required at all. The stress-ng 
> termination signal sequence is SIGALRMx4, SIGTERM then SIGKILL with a 
> small time between signals. I could change that to SIGALRM then 
> SIGKILL and lower the time interval to see if I can get a SIGALRM 
> being processed whilst SIGKILL is delivered without swapping having 
> taken place.
I have had no success yet creating a separate test case to reproduce 
these circumstances.
> I'd like to try and find the cause of the swapping bugs so I'll also 
> continue with the existing test case.

I had spent some time building a new test bed with the latest hurd code 
using rumpdisk only to find that swapping fails very quickly using this 
setup. I've not investigated precisely why yet but that seems to me an 
important area to investigate and fix. I'll look at this once I have 
resolved the current problem.

I have had some minor progress however on the original test scenario. It 
seems that it is not pageout/pagein that is causing the problem here but 
rather an assertion that is occurring within ext2fs. In my testcase the 
assertion actually deadlocks with mutex locking issues causing all 
message servicing threads within ext2fs to become blocked awaiting a 
mutex. That there are no receivers available for 
memory_object_data_request() messages to be accepted from the kernel is 
the reason that so many kernel threads become stuck in vm_fault_continue.

Anyway, the assertion in question occurs within 
glibc-2.41/sysdeps/mach/hurd/mig-reply.c:__mig_dealloc_reply_port(). 
Adding some debug to this code has shown that the existing thread local 
storage had a different port to that expected by the message header but 
that both are non-zero. I would think that it is more likely that the 
thread local storage is invalid rather than the message header but that 
remains to be seen.

I am posting this message really to see if anyone is aware of issues in 
the thread local storage code and whether anything similar has arisen 
before or indeed any other immediate thoughts as to where the root of 
the problem might be. Having looked at the 'tls' code I can see that any 
attempts to trace/debug this might be very difficult.

Regards,

Mike.