Re: [PATCH] kfence: Avoid stalling work queue task without allocations

From: Paul E. McKenney
Date: Thu Nov 19 2020 - 13:49:01 EST

On Thu, Nov 19, 2020 at 06:02:59PM +0100, Marco Elver wrote:
> On Thu, Nov 19, 2020 at 07:14AM -0800, Paul E. McKenney wrote:
> > On Thu, Nov 19, 2020 at 01:53:57PM +0100, Marco Elver wrote:
> > > On Wed, Nov 18, 2020 at 03:38PM -0800, Paul E. McKenney wrote:
> > > > On Wed, Nov 18, 2020 at 11:56:21PM +0100, Marco Elver wrote:
> > > > > [...]
> > > > > I think I figured out one piece of the puzzle. Bisection keeps pointing
> > > > > me at some -rcu merge commit, which kept throwing me off. Nor did it
> > > > > help that reproduction is a bit flaky. However, I think there are 2
> > > > > independent problems, but the manifestation of 1 problem triggers the
> > > > > 2nd problem:
> > > > >
> > > > > 1. problem: slowed forward progress (workqueue lockup / RCU stall reports)
> > > > >
> > > > > 2. problem: DEADLOCK which causes complete system lockup
> > > > >
> > > > > | ...
> > > > > | CPU0
> > > > > | ----
> > > > > | lock(rcu_node_0);
> > > > > | <Interrupt>
> > > > > | lock(rcu_node_0);
> > > > > |
> > > > > | *** DEADLOCK ***
> > > > > |
> > > > > | 1 lock held by event_benchmark/105:
> > > > > | #0: ffffbb6e0b804458 (rcu_node_0){?.-.}-{2:2}, at: print_other_cpu_stall kernel/rcu/tree_stall.h:493 [inline]
> > > > > | #0: ffffbb6e0b804458 (rcu_node_0){?.-.}-{2:2}, at: check_cpu_stall kernel/rcu/tree_stall.h:652 [inline]
> > > > > | #0: ffffbb6e0b804458 (rcu_node_0){?.-.}-{2:2}, at: rcu_pending kernel/rcu/tree.c:3752 [inline]
> > > > > | #0: ffffbb6e0b804458 (rcu_node_0){?.-.}-{2:2}, at: rcu_sched_clock_irq+0x428/0xd40 kernel/rcu/tree.c:2581
> > > > > | ...
> > > > >
> > > > > Problem 2 can with reasonable confidence (5 trials) be fixed by reverting:
> > > > >
> > > > > rcu: Don't invoke try_invoke_on_locked_down_task() with irqs disabled
> > > > >
> > > > > At which point the system always boots to user space -- albeit with a
> > > > > bunch of warnings still (attached). The supposed "good" version doesn't
> > > > > end up with all those warnings deterministically, so I couldn't say if
> > > > > the warnings are expected due to recent changes or not (Arm64 QEMU
> > > > > emulation, 1 CPU, and lots of debugging tools on).
> > > > >
> > > > > Does any of that make sense?
> > > >
> > > > Marco, it makes all too much sense! :-/
> > > >
> > > > Does the patch below help?
> > > >
> > > > Thanx, Paul
> > > >
> > > > ------------------------------------------------------------------------
> > > >
> > > > commit 444ef3bbd0f243b912fdfd51f326704f8ee872bf
> > > > Author: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > > > Date: Sat Aug 29 10:22:24 2020 -0700
> > > >
> > > > sched/core: Allow try_invoke_on_locked_down_task() with irqs disabled
> > >
> > > My assumption is that this is a replacement for "rcu: Don't invoke
> > > try_invoke_on_locked_down_task() with irqs disabled", right?
> >
> > Hmmm... It was actually intended to be in addition.
> >
> > > That seems to have the same result (same test setup) as only reverting
> > > "rcu: Don't invoke..." does: still results in a bunch of workqueue
> > > lockup warnings and RCU stall warnings, but boots to user space. I
> > > attached a log. If the warnings are expected (are they?), then it looks
> > > fine to me.
> >
> > No, they are not at all expected, but might be a different symptom
> > of the original problem. Please see below.
> >
> > > (And just in case: with "rcu: Don't invoke..." and "sched/core:
> > > Allow..." both applied I still get DEADLOCKs -- but that's probably
> > > expected.)
> >
> > As noted earlier, it is a surprise. Could you please send me the
> > console output?
>
> I've attached the output of a run with both commits applied.

Got it, thank you!

> > > Testing all events: OK
> > > hrtimer: interrupt took 17120368 ns
> > > Running tests again, along with the function tracer
> > > Running tests on all trace events:
> > > Testing all events:
> > > BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 12s!
> > > Showing busy workqueues and worker pools:
> > > workqueue events: flags=0x0
> > > pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/256 refcnt=2
> > > pending: vmstat_shepherd
> > > BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 17s!
> > > Showing busy workqueues and worker pools:
> > > workqueue events: flags=0x0
> > > pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/256 refcnt=2
> > > pending: vmstat_shepherd
> > > workqueue events_power_efficient: flags=0x82
> > > pwq 2: cpus=0 flags=0x4 nice=0 active=1/256 refcnt=3
> > > pending: neigh_periodic_work
> > > ------------[ cut here ]------------
> > > WARNING: CPU: 0 PID: 1 at kernel/rcu/tree_stall.h:758 rcu_check_gp_start_stall kernel/rcu/tree_stall.h:750 [inline]
> > > WARNING: CPU: 0 PID: 1 at kernel/rcu/tree_stall.h:758 rcu_check_gp_start_stall.isra.0+0x14c/0x210 kernel/rcu/tree_stall.h:711
> >
> > I have different line numbering,
>
> This is still using next-20201110. I'll rerun with latest -next as well.

No problem, as it looks like next-20201105 is a reasonable approximation.

> > but the only warning that I see in this
> > function is the one complaining that RCU has been ignoring a request to
> > start a grace period for too long. This usually happens because the RCU
> > grace-period kthread (named "rcu_preempt" in your case, but can also be
> > named "rcu_sched") is being prevented from running, but can be caused
> > by other things as well.
> >
> > > Modules linked in:
> > > CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.10.0-rc3-next-20201110-00003-g920304642405-dirty #30
> > > Hardware name: linux,dummy-virt (DT)
> > > pstate: 20000085 (nzCv daIf -PAN -UAO -TCO BTYPE=--)
> > > pc : rcu_check_gp_start_stall kernel/rcu/tree_stall.h:750 [inline]
> > > pc : rcu_check_gp_start_stall.isra.0+0x14c/0x210 kernel/rcu/tree_stall.h:711
> > > lr : __xchg_mb arch/arm64/include/asm/cmpxchg.h:88 [inline]
> > > lr : atomic_xchg include/asm-generic/atomic-instrumented.h:615 [inline]
> > > lr : rcu_check_gp_start_stall kernel/rcu/tree_stall.h:751 [inline]
> > > lr : rcu_check_gp_start_stall.isra.0+0x148/0x210 kernel/rcu/tree_stall.h:711
> >
> > Two program counters and four link registers? Awesome! ;-)
>
> Ah I'm using syzkaller's symbolizer, which duplicates lines if there was
> an inline function (remove all the "[inline]" and it should make sense,
> but the "[inline]" tell you the actual line). Obviously for things like
> this it's a bit unintuitive. :-)

Very useful, though, and a big THANK YOU to those who made it happen!

> > > sp : ffff800010003d20
> > > x29: ffff800010003d20 x28: ffff274ac3a10000
> > > x27: 0000000000000000 x26: ffff274b3dbe72d8
> > > x25: ffffbcb867722000 x24: 0000000000000000
> > > x23: 0000000000000000 x22: ffffbcb8681d1260
> > > x21: ffffbcb86735b000 x20: ffffbcb867404440
> > > x19: ffffbcb867404440 x18: 0000000000000123
> > > x17: ffffbcb865d400f0 x16: 0000000000000002
> > > x15: 0000000000000002 x14: 0000000000000000
> > > x13: 003d090000000000 x12: 00001e8480000000
> > > x11: ffffbcb867958980 x10: ffff800010003cf0
> > > x9 : ffffbcb864f4b7c8 x8 : 0000000000000080
> > > x7 : 0000000000000026 x6 : ffffbcb86774e4c0
> > > x5 : 0000000000000000 x4 : 00000000d4001f4b
> > > x3 : 0000000000000000 x2 : 0000000000000000
> > > x1 : 0000000000000001 x0 : 0000000000000000
> > > Call trace:
> > > rcu_check_gp_start_stall kernel/rcu/tree_stall.h:750 [inline]
> > > rcu_check_gp_start_stall.isra.0+0x14c/0x210 kernel/rcu/tree_stall.h:711
> > > rcu_core+0x168/0x9e0 kernel/rcu/tree.c:2719
> > > rcu_core_si+0x18/0x28 kernel/rcu/tree.c:2737
> >
> > The RCU_SOFTIRQ handler is causing this checking to occur, for whatever
> > that is worth.
> >
> > > __do_softirq+0x188/0x6b4 kernel/softirq.c:298
> > > do_softirq_own_stack include/linux/interrupt.h:568 [inline]
> > > invoke_softirq kernel/softirq.c:393 [inline]
> > > __irq_exit_rcu kernel/softirq.c:423 [inline]
> > > irq_exit+0x1cc/0x1e0 kernel/softirq.c:447
> > > __handle_domain_irq+0xb4/0x130 kernel/irq/irqdesc.c:690
> > > handle_domain_irq include/linux/irqdesc.h:170 [inline]
> > > gic_handle_irq+0x70/0x108 drivers/irqchip/irq-gic.c:370
> > > el1_irq+0xc0/0x180 arch/arm64/kernel/entry.S:651
> > > arch_local_irq_restore+0x8/0x10 arch/arm64/include/asm/irqflags.h:124
> > > release_probes kernel/tracepoint.c:113 [inline]
> > > tracepoint_remove_func kernel/tracepoint.c:315 [inline]
> > > tracepoint_probe_unregister+0x220/0x378 kernel/tracepoint.c:382
> > > trace_event_reg+0x58/0x150 kernel/trace/trace_events.c:298
> > > __ftrace_event_enable_disable+0x424/0x608 kernel/trace/trace_events.c:412
> > > ftrace_event_enable_disable kernel/trace/trace_events.c:495 [inline]
> > > __ftrace_set_clr_event_nolock+0x120/0x180 kernel/trace/trace_events.c:811
> > > __ftrace_set_clr_event+0x60/0x90 kernel/trace/trace_events.c:833
> > > event_trace_self_tests+0xd4/0x114 kernel/trace/trace_events.c:3661
> > > event_trace_self_test_with_function kernel/trace/trace_events.c:3734 [inline]
> > > event_trace_self_tests_init+0x88/0xa8 kernel/trace/trace_events.c:3747
> > > do_one_initcall+0xa4/0x500 init/main.c:1212
> > > do_initcall_level init/main.c:1285 [inline]
> > > do_initcalls init/main.c:1301 [inline]
> > > do_basic_setup init/main.c:1321 [inline]
> > > kernel_init_freeable+0x344/0x3c4 init/main.c:1521
> > > kernel_init+0x20/0x16c init/main.c:1410
> > > ret_from_fork+0x10/0x34 arch/arm64/kernel/entry.S:961
> > > irq event stamp: 3274113
> > > hardirqs last enabled at (3274112): [<ffffbcb864f8aee4>] rcu_core+0x974/0x9e0 kernel/rcu/tree.c:2716
> > > hardirqs last disabled at (3274113): [<ffffbcb866233bf0>] __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:108 [inline]
> > > hardirqs last disabled at (3274113): [<ffffbcb866233bf0>] _raw_spin_lock_irqsave+0xb8/0x14c kernel/locking/spinlock.c:159
> > > softirqs last enabled at (3272576): [<ffffbcb864e10b80>] __do_softirq+0x630/0x6b4 kernel/softirq.c:325
> > > softirqs last disabled at (3274101): [<ffffbcb864ec6c54>] do_softirq_own_stack include/linux/interrupt.h:568 [inline]
> > > softirqs last disabled at (3274101): [<ffffbcb864ec6c54>] invoke_softirq kernel/softirq.c:393 [inline]
> > > softirqs last disabled at (3274101): [<ffffbcb864ec6c54>] __irq_exit_rcu kernel/softirq.c:423 [inline]
> > > softirqs last disabled at (3274101): [<ffffbcb864ec6c54>] irq_exit+0x1cc/0x1e0 kernel/softirq.c:447
> > > ---[ end trace 902768efebf5a607 ]---
> > > rcu: rcu_preempt: wait state: RCU_GP_WAIT_GPS(1) ->state: 0x0 delta ->gp_activity 4452 ->gp_req_activity 3848 ->gp_wake_time 3848 ->gp_wake_seq 2696 ->gp_seq 2696 ->gp_seq_needed 2700 ->gp_flags 0x1
> >
> > The last thing that RCU's grace-period kthread did was to go to sleep
> > waiting for a grace-period request (RCU_GP_WAIT_GPS).
> >
> > > rcu: rcu_node 0:0 ->gp_seq 2696 ->gp_seq_needed 2700
> > > rcu: RCU callbacks invoked since boot: 2583
> > > rcu_tasks: RTGS_WAIT_CBS(11) since 567120 g:1 i:0/0 k.
> > > rcu_tasks_rude: RTGS_WAIT_CBS(11) since 567155 g:1 i:0/1 k.
> > > rcu_tasks_trace: RTGS_INIT(0) since 4295464549 g:0 i:0/0 k. N0 h:0/0/0
> > > rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
> > > (detected by 0, t=3752 jiffies, g=2705, q=8)
> > > rcu: All QSes seen, last rcu_preempt kthread activity 557 (4295471128-4295470571), jiffies_till_next_fqs=1, root ->qsmask 0x0
> > > rcu: rcu_preempt kthread starved for 557 jiffies! g2705 f0x2 RCU_GP_CLEANUP(7) ->state=0x0 ->cpu=0
> >
> > And here we see that RCU's grace-period kthread has in fact been starved.
> >
> > This kthread is now in RCU_GP_CLEANUP, perhaps because of the wakeup that is
> > sent in rcu_check_gp_kthread_starvation().
> >
> > My current guess is that this is a consequence of the earlier failures,
> > but who knows?
>
> I can try bisection again, or reverting some commits that might be
> suspicious? But we'd need some selection of suspicious commits.

The report claims that one of the rcu_node ->lock fields is held
with interrupts enabled, which would indeed be bad. Except that all
of the stack traces that it shows have these locks held within the
scheduling-clock interrupt handler. Now with the "rcu: Don't invoke
try_invoke_on_locked_down_task() with irqs disabled" but without the
"sched/core: Allow try_invoke_on_locked_down_task() with irqs disabled"
commit, I understand why. With both, I don't see how this happens.

At this point, I am reduced to adding lockdep_assert_irqs_disabled()
calls at various points in that code, as shown in the patch below.

At this point, I would guess that your first priority would be the
initial bug rather than this following issue, but you never know, this
might well help diagnose the initial bug.

Thanx, Paul

------------------------------------------------------------------------

commit ccedf00693ef60f7c06d23490fc41bb60dd43dc3
Author: Paul E. McKenney <paulmck@xxxxxxxxxx>
Date: Thu Nov 19 10:13:06 2020 -0800

rcu: Add lockdep_assert_irqs_disabled() to rcu_sched_clock_irq() and callees

This commit adds a number of lockdep_assert_irqs_disabled() calls
to rcu_sched_clock_irq() and a number of the functions that it calls.
The point of this is to help track down a situation where lockdep appears
to be insisting that interrupts are enabled within these functions, which
should only ever be invoked from the scheduling-clock interrupt handler.

Link: https://lore.kernel.org/lkml/20201111133813.GA81547@xxxxxxxxxxxxxxxx/
Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 2b3274c..1d956f9 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2580,6 +2580,7 @@ static void rcu_do_batch(struct rcu_data *rdp)
void rcu_sched_clock_irq(int user)
{
trace_rcu_utilization(TPS("Start scheduler-tick"));
+ lockdep_assert_irqs_disabled();
raw_cpu_inc(rcu_data.ticks_this_gp);
/* The load-acquire pairs with the store-release setting to true. */
if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
@@ -2593,6 +2594,7 @@ void rcu_sched_clock_irq(int user)
rcu_flavor_sched_clock_irq(user);
if (rcu_pending(user))
invoke_rcu_core();
+ lockdep_assert_irqs_disabled();

trace_rcu_utilization(TPS("End scheduler-tick"));
}
@@ -3761,6 +3763,8 @@ static int rcu_pending(int user)
struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
struct rcu_node *rnp = rdp->mynode;

+ lockdep_assert_irqs_disabled();
+
/* Check for CPU stalls, if enabled. */
check_cpu_stall(rdp);

diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
index 642ebd6..f7aa41c 100644
--- a/kernel/rcu/tree_plugin.h
+++ b/kernel/rcu/tree_plugin.h
@@ -682,6 +682,7 @@ static void rcu_flavor_sched_clock_irq(int user)
{
struct task_struct *t = current;

+ lockdep_assert_irqs_disabled();
if (user || rcu_is_cpu_rrupt_from_idle()) {
rcu_note_voluntary_context_switch(current);
}
diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
index 4e3aecd..f276d8e 100644
--- a/kernel/rcu/tree_stall.h
+++ b/kernel/rcu/tree_stall.h
@@ -267,6 +267,7 @@ static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
struct task_struct *t;
struct task_struct *ts[8];

+ lockdep_assert_irqs_disabled();
if (!rcu_preempt_blocked_readers_cgp(rnp))
return 0;
pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
@@ -291,6 +292,7 @@ static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
".q"[rscr.rs.b.need_qs],
".e"[rscr.rs.b.exp_hint],
".l"[rscr.on_blkd_list]);
+ lockdep_assert_irqs_disabled();
put_task_struct(t);
ndetected++;
}
@@ -527,6 +529,8 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
struct rcu_node *rnp;
long totqlen = 0;

+ lockdep_assert_irqs_disabled();
+
/* Kick and suppress, if so configured. */
rcu_stall_kick_kthreads();
if (rcu_stall_is_suppressed())
@@ -548,6 +552,7 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
}
}
ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
+ lockdep_assert_irqs_disabled();
}

for_each_possible_cpu(cpu)
@@ -594,6 +599,8 @@ static void print_cpu_stall(unsigned long gps)
struct rcu_node *rnp = rcu_get_root();
long totqlen = 0;

+ lockdep_assert_irqs_disabled();
+
/* Kick and suppress, if so configured. */
rcu_stall_kick_kthreads();
if (rcu_stall_is_suppressed())
@@ -649,6 +656,7 @@ static void check_cpu_stall(struct rcu_data *rdp)
unsigned long js;
struct rcu_node *rnp;

+ lockdep_assert_irqs_disabled();
if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
!rcu_gp_in_progress())
return;