Re: [PATCH] stacktrace: do not trace user stack for user_worker tasks

[Steven Rostedt]

[
  Adding Peter Zijlstra as he has been telling me to test against
  PF_KTHREAD instead of current->mm to tell if it is a kernel thread.
  But that seems to not be enough!
]

On Wed, 25 Jun 2025 10:23:28 -0600
Jens Axboe <axboe@xxxxxxxxx> wrote:

> On 6/24/25 11:07 AM, Steven Rostedt wrote:
> > On Mon, 23 Jun 2025 19:59:11 +0800
> > Jiazi Li <jqqlijiazi@xxxxxxxxx> wrote:
> >   
> >> Tasks with PF_USER_WORKER flag also only run in kernel space,
> >> so do not trace user stack for these tasks.  
> > 
> > What exactly is the difference between PF_KTHREAD and PF_USER_WORKER?  
> 
> One is a kernel thread (eg no mm, etc), the other is basically a user
> thread. None of them exit to userspace, that's basically the only
> thing they have in common.

Was it ever in user space? Because exiting isn't the issue for getting
a user space stack. If it never was in user space than sure, there's no
reason to look at the user space stack.

> 
> > Has all the locations that test for PF_KTHREAD been audited to make
> > sure that PF_USER_WORKER isn't also needed?  
> 
> I did when adding it, to the best of my knowledge. But there certainly
> could still be gaps. Sometimes not easy to see why code checks for
> PF_KTHREAD in the first place.
> 
> > I'm working on other code that needs to differentiate between user
> > tasks and kernel tasks, and having to have multiple flags to test is
> > becoming quite a burden.  
> 
> None of them are user tasks, but PF_USER_WORKER does look like a
> user thread and acts like one, except it wasn't created by eg
> pthread_create() and it never returns to userspace. When it's done,
> it's simply reaped.
> 

I'm assuming that it also never was in user space, which is where we
don't want to do any user space stack trace.

This looks like more rationale for having a kernel_task() user_task()
helper functions:

  https://lore.kernel.org/linux-trace-kernel/20250425204120.639530125@xxxxxxxxxxx/

Where one returns true for both PF_KERNEL and PF_USER_WORKER and the
other returns false.

-- Steve