Re: [PATCH] kgdb: Avoid suspicious RCU usage warning

From: Doug Anderson
Date: Tue Jun 02 2020 - 18:56:49 EST


Hi,

On Mon, Jun 1, 2020 at 9:20 AM Daniel Thompson
<daniel.thompson@xxxxxxxxxx> wrote:
>
> On Wed, May 27, 2020 at 05:02:27PM -0700, Doug Anderson wrote:
> > Hi,
> >
> > On Tue, May 19, 2020 at 3:41 AM Daniel Thompson
> > <daniel.thompson@xxxxxxxxxx> wrote:
> > >
> > > On Thu, May 07, 2020 at 03:53:58PM -0700, Douglas Anderson wrote:
> > > > At times when I'm using kgdb I see a splat on my console about
> > > > suspicious RCU usage. I managed to come up with a case that could
> > > > reproduce this that looked like this:
> > > >
> > > > WARNING: suspicious RCU usage
> > > > 5.7.0-rc4+ #609 Not tainted
> > > > -----------------------------
> > > > kernel/pid.c:395 find_task_by_pid_ns() needs rcu_read_lock() protection!
> > > >
> > > > other info that might help us debug this:
> > > >
> > > > rcu_scheduler_active = 2, debug_locks = 1
> > > > 3 locks held by swapper/0/1:
> > > > #0: ffffff81b6b8e988 (&dev->mutex){....}-{3:3}, at: __device_attach+0x40/0x13c
> > > > #1: ffffffd01109e9e8 (dbg_master_lock){....}-{2:2}, at: kgdb_cpu_enter+0x20c/0x7ac
> > > > #2: ffffffd01109ea90 (dbg_slave_lock){....}-{2:2}, at: kgdb_cpu_enter+0x3ec/0x7ac
> > > >
> > > > stack backtrace:
> > > > CPU: 7 PID: 1 Comm: swapper/0 Not tainted 5.7.0-rc4+ #609
> > > > Hardware name: Google Cheza (rev3+) (DT)
> > > > Call trace:
> > > > dump_backtrace+0x0/0x1b8
> > > > show_stack+0x1c/0x24
> > > > dump_stack+0xd4/0x134
> > > > lockdep_rcu_suspicious+0xf0/0x100
> > > > find_task_by_pid_ns+0x5c/0x80
> > > > getthread+0x8c/0xb0
> > > > gdb_serial_stub+0x9d4/0xd04
> > > > kgdb_cpu_enter+0x284/0x7ac
> > > > kgdb_handle_exception+0x174/0x20c
> > > > kgdb_brk_fn+0x24/0x30
> > > > call_break_hook+0x6c/0x7c
> > > > brk_handler+0x20/0x5c
> > > > do_debug_exception+0x1c8/0x22c
> > > > el1_sync_handler+0x3c/0xe4
> > > > el1_sync+0x7c/0x100
> > > > rpmh_rsc_probe+0x38/0x420
> > > > platform_drv_probe+0x94/0xb4
> > > > really_probe+0x134/0x300
> > > > driver_probe_device+0x68/0x100
> > > > __device_attach_driver+0x90/0xa8
> > > > bus_for_each_drv+0x84/0xcc
> > > > __device_attach+0xb4/0x13c
> > > > device_initial_probe+0x18/0x20
> > > > bus_probe_device+0x38/0x98
> > > > device_add+0x38c/0x420
> > > >
> > > > If I understand properly we should just be able to blanket kgdb under
> > > > one big RCU read lock and the problem should go away. We'll add it to
> > > > the beast-of-a-function known as kgdb_cpu_enter().
> > > >
> > > > With this I no longer get any splats and things seem to work fine.
> > > >
> > > > Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
> > >
> > > In principle this looks OK but I'm curious why we don't cuddle these
> > > calls up to the local interrupt locking (and also whether we want to
> > > keep hold of the lock during stepping). If nothing else that would make
> > > review easier.
> >
> > It probably wouldn't hurt to keep hold of the lock during single
> > stepping but I don't think there's any real reason we'd want to.
> > Specifically the only real reason we're calling rcu_read_lock() is to
> > avoid the warning. Since we're a stop-the-world debugger it's not
> > like something else could be messing with state at the same time.
> >
> > I'm looking at the whole function though and I don't really understand
> > all the comments about interrupts being restored by the 'trap return'
> > code, do you?
>
> Not as much as I'd like but, in principle at least, I think the trap
> handler is written to make it an architecture decision to whether it
> will be entered with interrupt disabled or not (which would in turn
> depend on how the architecture manages exception stacks and what else
> it uses the breakpoint machinery for).
>
>
> > Specifically: as far as I can tell we _always_ restore
> > interrupts when exiting the function. There are only two return
> > statements and both have "local_irq_restore(flags);" right before
> > them. We never modify the flags directly and the one other usage of
> > "flags" is effectively the statement "local_irq_restore(flags);
> > local_irq_save(flags);" which will, I guess, allow any interrupts that
> > were already pending to take place. Are you saying that you want me
> > to match that and do a "rcu_read_unlock(); rcu_read_lock()" there?
> >
> > If I understand things correctly (and there's maybe a better chance
> > after I read Wei Li's recent patches) the disabling of IRQs for single
> > stepping happens in a different way. It looks like we update the
> > "struct pt_regs" of the task we're stepping so that when we exit kgdb
> > and start running the task again that the interrupts are off. That
> > seems reasonable to me and this function has nothing to do with it.
> >
> > ...and further confusion on my part: does the whole saving / restoring
> > of interrupts in kgdb_cpu_enter() make any sense anyway? Is this
> > function ever called from a context that's not an interrupt context?
> > How do we get the pt_regs in that case? Just for fun, I tried doing
> > this:
> >
> > local_irq_save(flags);
> > + if (!arch_irqs_disabled_flags(flags))
> > + pr_warn("I was wrong\n");
> >
> > ...and I never saw "I was wrong" on my system. Maybe it matters for
> > something not arm64? ...or, maybe, this is from when kgdb worked in a
> > completely different way?
>
> I'm not yet in a position to test kgdb on all architectures. Mostly
> this is because it can be hard to discover suitable qemu stanzas to
> launch an emulator... and in one case because buildroot doesn't support
> cross-gdb yet. Anyhow, right now, it is difficult to tour the niche
> architectures to give a definitive answer.

OK, fair enough. I guess I just really don't understand how it could
work if an architecture didn't do this, but perhaps it would be clear
if there was code demonstrating it. ;-)


> > In general I made my patch by:
> > * Calling rcu_read_lock() at the start of the function.
> > * Calling rcu_read_unlock() right before all 2 of the "return" calls of
> > the function.
> >
> > ...I was hoping that would actually make it easier to reason about
> > even if the function is a beast.
> >
> >
> > Hopefully the above makes sense. I wouldn't rule out me just being
> > utterly confused, but I _think_ I reasoned through it all. ;-) If it
> > all makes sense, I'm inclined to:
> >
> > 1. Leave my patch the way it is.
>
> I'm still a little reluctant on this.
>
> It's not that I think the current patch is buggy, more that I view the
> interrupt locks as the place in the function that we start to "assert"
> what our execution context is going to be. Seeing all the entry/exit
> logic in one place makes reasoning easier and should make the eventual
> refactoring of "the beast" a little easier too.

OK, I give in. Posted v2 doing it how I think you want it.


> > 2. Perhaps remove the whole irq saving / restoring in kgdb_cpu_enter().
>
> Are you feeling lucky?
>
> I think there will come a time when bravery is called for but I'd rather
> see this as part of a bigger rewrite instead of a single high risk
> change.

Hrm, maybe. I guess it depends on whether we want to take baby steps
there or try to do it all at once. If we take baby steps we will
occasionally fall down but we'll slowly start getting things cleaned
up. If we wait for a full rewrite then we might be waiting for a long
time. It'll also be harder to figure out which of the big changes in
the major rewrite broken someone. ...or if the major rewrite comes in
20 small/bisectable patches it may be hard to revert patch 2 out of 20
if the future patches all build upon it. If we do one small high-risk
change and then wait before building upon it then it'll be easy for
someone to bisect and then yell for a revert.


-Doug