[PATCH] ftrace: Fixup lockdep assert held of text_mutex

Steven Rostedt rostedt at goodmis.org
Thu Aug 13 11:37:43 EDT 2020

On Wed, 12 Aug 2020 22:13:19 -0700 (PDT)
Palmer Dabbelt <palmer at dabbelt.com> wrote:

> Sorry, I'm not really sure what's going on here.  I'm not really seeing code
> that matches this in our port right now, so maybe this is aginst some other
> tree?  If it's the RISC-V kprobes patch set then I was hoping to take a look at
> that tomorrow (or I guess a bit earlier this week, but I had some surprise work
> stuff to do).  IIRC there were a handful of races in the last patch set I saw,
> but it's been a while so I don't remember for sure.
> That said, I certainly wouldn't be surprised if there's a locking bug in our
> ftrace stuff.  It'd be way easier for me to figure out what's going on if you
> have a concrete suggestion as to how to fix the issues -- even if it's just a
> workaround.

The issue is actually quite basic.

ftrace_init_nop() is called quite early in boot up and never called
again. It's called before SMP is set up, so it's on a single CPU, and
no worries about synchronization with other CPUs is needed.

On x86, it is called before text_poke() is initialized (which is used
to synchronize code updates across CPUs), and thus can't be called.
There's a "text_poke_early()" that is used instead, which is basically
just a memcpy().

Now, if ftrace_init_nop() is not defined by the architecture, it is a
simple call to ftrace_make_nop(), which is also used to disable ftrace

The issue is that we have the following path on riscv:


Boom! text_mutex is not held, and lockdep complains.

The difference between ftrace_make_nop() being called by
ftrace_init_nop() and being called later to disable function tracing is
that the latter will have:


and the former will not have those called.

On x86, we handle the two different cases with:

static int ftrace_poke_late = 0;

int ftrace_arch_code_modify_prepare(void)
	ftrace_poke_late = 1;
	return 0;

int ftrace_arch_code_modify_post_process(void)
	ftrace_poke_late = 0;

Although, the post_process() probably doesn't even need to set
ftrace_poke_late back to zero.

Then in ftrace_make_nop(), we have:

      if (ftrace_poke_late)
        text_poke_queue(...); // this checks if text_mutex is held
        text_poke_early(...); // is basically just memcpy, no test on text_mutex.

The two solutions for riscv, is either to implement the same thing as
above, or you can create your own ftrace_init_nop() to take the
text_mutex before calling ftrace_make_nop(), and that should work too.

-- Steve

More information about the linux-riscv mailing list