Date:   Wed, 24 May 2017 18:37:35 +0900
From:   Masami Hiramatsu <mhiramat@...nel.org>
To:     Steven Rostedt <rostedt@...dmis.org>
Cc:     "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
        Ingo Molnar <mingo@...nel.org>, linux-kernel@...r.kernel.org,
        Masami Hiramatsu <mhiramat@...nel.org>
Subject: Re: Use case for TASKS_RCU

Sorry, I missed this thread,

On Tue, 16 May 2017 09:07:08 -0400
Steven Rostedt <rostedt@...dmis.org> wrote:

> On Tue, 16 May 2017 05:23:54 -0700
> "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com> wrote:
> 
> > On Tue, May 16, 2017 at 08:22:33AM +0200, Ingo Molnar wrote:
> > > 
> > > * Paul E. McKenney <paulmck@...ux.vnet.ibm.com> wrote:
> > >   
> > > > Hello!
> > > > 
> > > > The question of the use case for TASKS_RCU came up, and here is my
> > > > understanding.  Steve will not be shy about correcting any misconceptions
> > > > I might have.  ;-)
> > > > 
> > > > The use case is to support freeing of trampolines used in tracing/probing
> > > > in CONFIG_PREEMPT=y kernels.  It is necessary to wait until any task
> > > > executing in the trampoline in question has left it, taking into account
> > > > that the trampoline's code might be interrupted and preempted.  However,
> > > > the code in the trampolines is guaranteed never to context switch.
> > > > 
> > > > Note that in CONFIG_PREEMPT=n kernels, synchronize_sched() suffices.
> > > > It is therefore tempting to think in terms of disabling preemption across
> > > > the trampolines, but there is apparently not enough room to accommodate
> > > > the needed preempt_disable() and preempt_enable() in the code invoking
> > > > the trampoline, and putting the preempt_disable() and preempt_enable()
> > > > in the trampoline itself fails because of the possibility of preemption
> > > > just before the preempt_disable() and just after the preempt_enable().
> > > > Similar reasoning rules out use of rcu_read_lock() and rcu_read_unlock().  
> > > 
> > > So how was this solved before TASKS_RCU? Also, nothing uses call_rcu_tasks() at 
> > > the moment, so it's hard for me to review its users. What am I missing?  
> > 
> > Before TASKS_RCU, the trampolines were just leaked when CONFIG_PREEMPT=y.
> 
> Actually, things simply were not implemented. This is why optimized
> kprobes is dependent on !CONFIG_PREEMPT. In fact, we can now optimize
> kprobes on CONFIG_PREEMPT with this utility. Right Masami?

Yes, I just haven't implemented it. OK, I'll use synchronize_rcu_tasks.

> With ftrace, perf and other "dynamic" users (where the ftrace_ops was
> created via a kmalloc), would not get the benefit of being called
> directly. They all needed to have their mcount/fentry's call a static
> trampoline that disabled preemption before calling the callback. This
> static trampoline is shared by all, so even if perf was the only
> callback for the function, it had to call this trampoline that iterated
> through all registered ftrace_ops to see which one had a callback for
> the given function.

For the optimized kprobes, it always jumps into dynamically allocated
trampoline, so we have no chance to disable preemption.

Thank you,

> 
> With this utility, perf not only gets the benefit of not having to use
> that static loop trampoline, it can even have its own trampoline
> created that doesn't even need to do the check if perf wants this
> function or not, as the only way the trampoline is called, is if perf
> wanted it.
> 
> > 
> > Current mainline kernel/trace/ftrace.c uses synchronize_rcu_tasks().
> > So yes, currently one user.
> > 
> 
> And the kpatch folks want to use it too.
> 
> -- Steve


-- 
Masami Hiramatsu <mhiramat@...nel.org>

Powered by blists - more mailing lists