[PATCH RESEND] mm: memcg/slab: fix panic in __free_slab() caused by premature memcg pointer release

[Roman Gushchin]

Karsten reported the following panic in __free_slab() happening on a s390x
machine:

349.361168Â Unable to handle kernel pointer dereference in virtual kernel address space
349.361210Â Failing address: 0000000000000000 TEID: 0000000000000483
349.361223Â Fault in home space mode while using kernel ASCE.
349.361240Â AS:00000000017d4007 R3:000000007fbd0007 S:000000007fbff000 P:000000000000003d
349.361340Â Oops: 0004 ilc:3 Ã#1Â PREEMPT SMP
349.361349Â Modules linked in: tcp_diag inet_diag xt_tcpudp ip6t_rpfilter ip6t_REJECT \
nf_reject_ipv6 ipt_REJECT nf_reject_ipv4 xt_conntrack ip6table_nat ip6table_mangle \
ip6table_raw ip6table_security iptable_at nf_nat
349.361436Â CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.3.0-05872-g6133e3e4bada-dirty #14
349.361445Â Hardware name: IBM 2964 NC9 702 (z/VM 6.4.0)
349.361450Â Krnl PSW : 0704d00180000000 00000000003cadb6 (__free_slab+0x686/0x6b0)
349.361464Â            R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:1 PM:0 RI:0 EA:3
349.361470Â Krnl GPRS: 00000000f3a32928 0000000000000000 000000007fbf5d00 000000000117c4b8
349.361475Â            0000000000000000 000000009e3291c1 0000000000000000 0000000000000000
349.361481Â            0000000000000003 0000000000000008 000000002b478b00 000003d080a97600
349.361481Â            0000000000000003 0000000000000008 000000002b478b00 000003d080a97600
349.361486Â            000000000117ba00 000003e000057db0 00000000003cabcc 000003e000057c78
349.361500Â Krnl Code: 00000000003cada6: e310a1400004        lg      %r1,320(%r10)
349.361500Â            00000000003cadac: c0e50046c286        brasl   %r14,ca32b8
349.361500Â           #00000000003cadb2: a7f4fe36            brc     15,3caa1e
349.361500Â           >00000000003cadb6: e32060800024        stg     %r2,128(%r6)
349.361500Â            00000000003cadbc: a7f4fd9e            brc     15,3ca8f8
349.361500Â            00000000003cadc0: c0e50046790c        brasl   %r14,c99fd8
349.361500Â            00000000003cadc6: a7f4fe2c            brc     15,3caa
349.361500Â            00000000003cadc6: a7f4fe2c            brc     15,3caa1e
349.361500Â            00000000003cadca: ecb1ffff00d9        aghik   %r11,%r1,-1
349.361619Â Call Trace:
349.361627Â (Ã<00000000003cabcc>Â __free_slab+0x49c/0x6b0)
349.361634Â  Ã<00000000001f5886>Â rcu_core+0x5a6/0x7e0
349.361643Â  Ã<0000000000ca2dea>Â __do_softirq+0xf2/0x5c0
349.361652Â  Ã<0000000000152644>Â irq_exit+0x104/0x130
349.361659Â  Ã<000000000010d222>Â do_IRQ+0x9a/0xf0
349.361667Â  Ã<0000000000ca2344>Â ext_int_handler+0x130/0x134
349.361674Â  Ã<0000000000103648>Â enabled_wait+0x58/0x128
349.361681Â (Ã<0000000000103634>Â enabled_wait+0x44/0x128)
349.361688Â  Ã<0000000000103b00>Â arch_cpu_idle+0x40/0x58
349.361695Â  Ã<0000000000ca0544>Â default_idle_call+0x3c/0x68
349.361704Â  Ã<000000000018eaa4>Â do_idle+0xec/0x1c0
349.361748Â  Ã<000000000018ee0e>Â cpu_startup_entry+0x36/0x40
349.361756Â  Ã<000000000122df34>Â arch_call_rest_init+0x5c/0x88
349.361761Â  Ã<0000000000000000>Â 0x0
349.361765Â INFO: lockdep is turned off.
349.361769Â Last Breaking-Event-Address:
349.361774Â  Ã<00000000003ca8f4>Â __free_slab+0x1c4/0x6b0
349.361781Â Kernel panic - not syncing: Fatal exception in interrupt

The kernel panics on an attempt to dereference the NULL memcg pointer.
When shutdown_cache() is called from the kmem_cache_destroy() context,
a memcg kmem_cache might have empty slab pages in a partial list,
which are still charged to the memory cgroup. These pages are released
by free_partial() at the beginning of shutdown_cache(): either
directly or by scheduling a RCU-delayed work (if the kmem_cache has
the SLAB_TYPESAFE_BY_RCU flag). The latter case is when the reported
panic can happen: memcg_unlink_cache() is called immediately after
shrinking partial lists, without waiting for scheduled RCU works.
It sets the kmem_cache->memcg_params.memcg pointer to NULL,
and the following attempt to dereference it by __free_slab()
from the RCU work context causes the panic.

To fix the issue, let's postpone the release of the memcg pointer
to destroy_memcg_params(). It's called from a separate work context
by slab_caches_to_rcu_destroy_workfn(), which contains a full RCU
barrier. This guarantees that all scheduled page release RCU works
will complete before the memcg pointer will be zeroed.

Big thanks for Karsten for the perfect report containing all necessary
information, his help with the analysis of the problem and testing
of the fix.

Fixes: fb2f2b0adb98 ("mm: memcg/slab: reparent memcg kmem_caches on cgroup removal")
Reported-by: Karsten Graul <kgraul@xxxxxxxxxxxxx>
Tested-by: Karsten Graul <kgraul@xxxxxxxxxxxxx>
Signed-off-by: Roman Gushchin <guro@xxxxxx>
Cc: Karsten Graul <kgraul@xxxxxxxxxxxxx>
Cc: Shakeel Butt <shakeelb@xxxxxxxxxx>
Cc: Vladimir Davydov <vdavydov.dev@xxxxxxxxx>
Cc: David Rientjes <rientjes@xxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx
---
 mm/slab_common.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/mm/slab_common.c b/mm/slab_common.c
index 0b94a37da531..8afa188f6e20 100644
--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -178,10 +178,13 @@ static int init_memcg_params(struct kmem_cache *s,
 
 static void destroy_memcg_params(struct kmem_cache *s)
 {
-	if (is_root_cache(s))
+	if (is_root_cache(s)) {
 		kvfree(rcu_access_pointer(s->memcg_params.memcg_caches));
-	else
+	} else {
+		mem_cgroup_put(s->memcg_params.memcg);
+		WRITE_ONCE(s->memcg_params.memcg, NULL);
 		percpu_ref_exit(&s->memcg_params.refcnt);
+	}
 }
 
 static void free_memcg_params(struct rcu_head *rcu)
@@ -253,8 +256,6 @@ static void memcg_unlink_cache(struct kmem_cache *s)
 	} else {
 		list_del(&s->memcg_params.children_node);
 		list_del(&s->memcg_params.kmem_caches_node);
-		mem_cgroup_put(s->memcg_params.memcg);
-		WRITE_ONCE(s->memcg_params.memcg, NULL);
 	}
 }
 #else
-- 
2.21.0