Re: [PATCH] riscv: ftrace: no need to acquire text_mutex when executed in stop_machine

[Palmer Dabbelt]

On Sun, 13 Mar 2022 00:07:11 PST (-0800), changbin.du@xxxxxxxxx wrote:
> On Thu, Mar 10, 2022 at 09:27:42AM -0500, Steven Rostedt wrote:
> > On Thu, 10 Mar 2022 12:54:54 +0800
> > Changbin Du <changbin.du@xxxxxxxxx> wrote:
> >
> > > It's safe to patch text segment in stop_machine. No race is possible here.
> > > Besides, there is a false positive for the lock assertion in
> > > patch_insn_write() since the lock is not held by cpu migration thread.
> > >
> > > So we actually don't need our ftrace_arch_code_modify_prepare/post(). And
> > > the lock assertion in patch_insn_write() should be removed to avoid
> > > producing lots of false positive warnings.
> > >
> > > Signed-off-by: Changbin Du <changbin.du@xxxxxxxxx>
> >
> > Ideally, RISC-V should try to get off of the stop_machine approach, and
> > move to the breakpoint modification.
> yes, that's a further step. I can feel a obvious stall to enable ftrace running
> in QEMU. (maybe qemu-riscv tcg is too slow...)

Looks like we've had this exact discussion before, even with exactly the 
same patch and people and everything.  I guess I dropped the ball here, 
I got so distracted trying to figure out those sequences to avoid 
stop_machine() that I forgot to clean up the patch to fix stop machine.  
I just sent out a v2 <20220322022331.32136-1-palmer@xxxxxxxxxxxx>.  That 
should at least fix the crash, we can deal with getting rid of 
stop_machine() later.