kmemleak: illegal RCU use assertion error

From: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
To: Catalin Marinas <catalin.marinas@arm.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>,
	linux-mm@kvack.org, linux-kernel@vger.kernel.org
Subject: kmemleak: illegal RCU use assertion error
Date: Mon, 2 Apr 2012 10:09:11 +0300	[thread overview]
Message-ID: <20120402070911.GB3464@swordfish> (raw)

Hello,

commit e5601400081651060a59bd1f45f2821bb8e97f95
Author: Paul E. McKenney <paul.mckenney@linaro.org>
Date:   Sat Jan 7 11:03:57 2012 -0800

    rcu: Simplify offline processing

    Move ->qsmaskinit and blkd_tasks[] manipulation to the CPU_DYING
    notifier.  This simplifies the code by eliminating a potential
    deadlock and by reducing the responsibilities of force_quiescent_state().
    Also rename functions to make their connection to the CPU-hotplug
    stages explicit.

    Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
    Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

introduced WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); to __call_rcu()
function, Paul also added cpu_offline checks to other routines (e.g. callbacks)
in later commits. It happens that kmemleak() triggers one of them.

During cpu core offline, kfree()->kmemleak_free()->put_object()-->__call_rcu() chain
for struct intel_shared_regs * is executed when no struct users left on this core -- 
all CPUs are dead or dying.

[ 4703.342462] CPU 3 is now offline
[ 4705.588116] ------------[ cut here ]------------
[ 4705.588129] WARNING: at kernel/rcutree.c:1823 __call_rcu+0x9d/0x1d2()
[..]
[ 4705.588196] Call Trace:
[ 4705.588207]  [<ffffffff81059a00>] ? synchronize_srcu+0x6/0x17
[ 4705.588215]  [<ffffffff8103364e>] warn_slowpath_common+0x83/0x9c
[ 4705.588223]  [<ffffffff8111e627>] ? get_object+0x31/0x31
[ 4705.588229]  [<ffffffff81033681>] warn_slowpath_null+0x1a/0x1c
[ 4705.588235]  [<ffffffff810af770>] __call_rcu+0x9d/0x1d2
[ 4705.588243]  [<ffffffff81013f52>] ? intel_pmu_cpu_dying+0x3b/0x5d
[ 4705.588249]  [<ffffffff810af8f1>] call_rcu_sched+0x17/0x19
[ 4705.588255]  [<ffffffff8111eb7e>] put_object+0x47/0x4b
[ 4705.588261]  [<ffffffff8111ed8b>] delete_object_full+0x2a/0x2e
[ 4705.588269]  [<ffffffff81491dc8>] kmemleak_free+0x26/0x45
[ 4705.588274]  [<ffffffff8111691f>] kfree+0x130/0x221
[ 4705.588280]  [<ffffffff81013f52>] intel_pmu_cpu_dying+0x3b/0x5d
[ 4705.588287]  [<ffffffff8149cb83>] x86_pmu_notifier+0xaf/0xb9
[ 4705.588296]  [<ffffffff814b0e9d>] notifier_call_chain+0xac/0xd9
[ 4705.588303]  [<ffffffff81059c9e>] __raw_notifier_call_chain+0xe/0x10
[ 4705.588309]  [<ffffffff810354ec>] __cpu_notify+0x20/0x37
[ 4705.588314]  [<ffffffff81035516>] cpu_notify+0x13/0x15
[ 4705.588320]  [<ffffffff81490fab>] take_cpu_down+0x28/0x2e
[ 4705.588326]  [<ffffffff8109ef7f>] stop_machine_cpu_stop+0x96/0xf1
[ 4705.588332]  [<ffffffff8109ece3>] cpu_stopper_thread+0xe3/0x183
[ 4705.588338]  [<ffffffff8109eee9>] ? queue_stop_cpus_work+0xd0/0xd0
[ 4705.588344]  [<ffffffff814ad382>] ? _raw_spin_unlock_irqrestore+0x47/0x65
[ 4705.588353]  [<ffffffff81087d0d>] ? trace_hardirqs_on_caller+0x119/0x175
[ 4705.588358]  [<ffffffff81087d76>] ? trace_hardirqs_on+0xd/0xf
[ 4705.588364]  [<ffffffff8109ec00>] ? cpu_stop_signal_done+0x2c/0x2c
[ 4705.588370]  [<ffffffff810544a9>] kthread+0x8b/0x93
[ 4705.588378]  [<ffffffff814b5f34>] kernel_thread_helper+0x4/0x10
[ 4705.588385]  [<ffffffff814ad7f0>] ? retint_restore_args+0x13/0x13
[ 4705.588391]  [<ffffffff8105441e>] ? __init_kthread_worker+0x5a/0x5a
[ 4705.588397]  [<ffffffff814b5f30>] ? gs_change+0x13/0x13
[ 4705.588400] ---[ end trace 720328982e35a713 ]---
[ 4705.588507] CPU 2 is now offline

My first solution was to return from delete_object() if object deallocation
performed on cpu_is_offline(smp_processor_id()), marking object with special
flag, say OBJECT_ORPHAN. And issue real object_delete() during scan (for example)
when we see OBJECT_ORPHAN object.  
That, however, requires to handle special case when cpu core offlined
for small period of time, leading to object insertion error in
create_object(), which either may be handled in 2 possible ways (assuming
that lookup_object() returned OBJECT_ORPHAN):
#1 delete orphaned object and retry with insertion (*)
#2 re-set existing orphan object

(*) performing delete_object() from within create_object() requires releasing
of held kmemleak and object locks, which is racy with other create_object() and
any possible scan() activities.

Yet I'm not exactly sure that option #2 is the correct one.
(I've kind of a patch [not properly tested, etc.] for #2 option).

	Sergey