lklfuse hang with stress-ng

[tavip]

stress-ng --class filesystem --all 0
stress-ng --class filesystem --seq 0 -v -t 60

[pscollins]

@tavip @opurdila How did you cause this, exactly? I suspect it comes from the same virtio issue as #64, which I've been trying to track down for the past few days --- it would be nice to have an additional testcase. Is this supposed to be run under lkl-hijack.sh? I get:

Unknown class: 'filesystem', available classes: cpu cpu-cache io interrupt memory network os scheduler vm

when I try to run it like that. I'm not sure if I've got the wrong version installed, etc.

IIRC you need a higher version, I've compiled the latest one instead of using the one installed on the system.

[Rondom]

This seems to occur when there is a moderate IO load. I can reproduce this both with stress-ng and with filebench. For example running the webserver workload without any changes, will trigger it. Reducing the number of threads to 3 will make it disappear, with 4 threads I can reliably reproduce it.

Or even by running the snippet below in shell and having the folder open in Nautilus (GNOME file manager) at the same time. Seems like only Nautilus creates enough IO to trigger it.

 while true; do; mkdir -p $MOUNT_POINT/foo; echo hi > $MOUNT_POINT/foo/bar; rm -rf $MOUNT_POINT/foo; done

What I see in the logs is that the number of IRQ skyrockets until all 64 IRQs are exhausted (lkl: syscall_thread: failed to allocate irq: -16). In between we get a few "syscall would block" messages.

So I am wondering what is the cause of this and what is happening. What is the expected behaviour when runnning out of IRQs (block, fail)? Why do we get those "syscall would block" messages?

[liuyuan10]

I don't use lklfuse so I can't comment on your application. The logs you showed is helpful:
"syscall would block" only happens when multiple host threads ask the default syscall thread to create a new syscall thread. It doesn't mean a error.

The "failed to allocate" error log might be the reason. As described in issue #147 , if you have ~64 host threads calling lkl syscalls, the error will happen.

To mitigate this, there are several ways in my mind:

1. call lkl_stop_syscall_thread() to free the irq. However, this won't help if you have 64 host threads calling lkl syscalls concurrently
2. Try PR lkl: Add support of host LKL syscall #160 . It doesn't require a irq for syscalls any more so this issue is solved. The PR may not be merged any time soon but it's pretty stable on network application.
3. setting auto_syscall_threads to false in arch/lkl/kernel/syscalls.c. Don't know if it still works. And it can cause severe performance issue because one blocking syscall will block all other threads and it causes deadlock easily.

[Rondom]

Thanks for your answer @liuyuan10!
I had already been following your pull request and it was the first thing I wanted to try the next morning. Indeed it fixes the issue.

The problem is that libfuse will call pthread_detach once it has reached more than 10 threads while creating new threads as needed. Right now there is no easy way to hook into that code without patching libfuse, as far as I can see. On could always write one's own FUSE loop function, though.

So indeed this is caused by #147, but we might still leave this open until we find a workaround in lklfuse.

[liuyuan10]

You are welcome. Would like to know if the patch works in your case. I need more test cases on that patch.

[Rondom]

See #192 for possible automatic cleanup making calling lkl_stop_syscall_thread() unnecessary.