Re: [PATCH RESEND] mm: memcg: synchronize objcg lists with a dedicated spinlock

From: Jeremy Linton
Date: Wed Feb 02 2022 - 10:58:31 EST


On 2/1/22 16:33, Roman Gushchin wrote:
Alexander reported a circular lock dependency revealed by the mmap1
ltp test:
LOCKDEP_CIRCULAR (suite: ltp, case: mtest06 (mmap1))
WARNING: possible circular locking dependency detected
5.17.0-20220113.rc0.git0.f2211f194038.300.fc35.s390x+debug #1 Not tainted
------------------------------------------------------
mmap1/202299 is trying to acquire lock:
00000001892c0188 (css_set_lock){..-.}-{2:2}, at: obj_cgroup_release+0x4a/0xe0
but task is already holding lock:
00000000ca3b3818 (&sighand->siglock){-.-.}-{2:2}, at: force_sig_info_to_task+0x38/0x180
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #1 (&sighand->siglock){-.-.}-{2:2}:
__lock_acquire+0x604/0xbd8
lock_acquire.part.0+0xe2/0x238
lock_acquire+0xb0/0x200
_raw_spin_lock_irqsave+0x6a/0xd8
__lock_task_sighand+0x90/0x190
cgroup_freeze_task+0x2e/0x90
cgroup_migrate_execute+0x11c/0x608
cgroup_update_dfl_csses+0x246/0x270
cgroup_subtree_control_write+0x238/0x518
kernfs_fop_write_iter+0x13e/0x1e0
new_sync_write+0x100/0x190
vfs_write+0x22c/0x2d8
ksys_write+0x6c/0xf8
__do_syscall+0x1da/0x208
system_call+0x82/0xb0
-> #0 (css_set_lock){..-.}-{2:2}:
check_prev_add+0xe0/0xed8
validate_chain+0x736/0xb20
__lock_acquire+0x604/0xbd8
lock_acquire.part.0+0xe2/0x238
lock_acquire+0xb0/0x200
_raw_spin_lock_irqsave+0x6a/0xd8
obj_cgroup_release+0x4a/0xe0
percpu_ref_put_many.constprop.0+0x150/0x168
drain_obj_stock+0x94/0xe8
refill_obj_stock+0x94/0x278
obj_cgroup_charge+0x164/0x1d8
kmem_cache_alloc+0xac/0x528
__sigqueue_alloc+0x150/0x308
__send_signal+0x260/0x550
send_signal+0x7e/0x348
force_sig_info_to_task+0x104/0x180
force_sig_fault+0x48/0x58
__do_pgm_check+0x120/0x1f0
pgm_check_handler+0x11e/0x180
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&sighand->siglock);
lock(css_set_lock);
lock(&sighand->siglock);
lock(css_set_lock);
*** DEADLOCK ***
2 locks held by mmap1/202299:
#0: 00000000ca3b3818 (&sighand->siglock){-.-.}-{2:2}, at: force_sig_info_to_task+0x38/0x180
#1: 00000001892ad560 (rcu_read_lock){....}-{1:2}, at: percpu_ref_put_many.constprop.0+0x0/0x168
stack backtrace:
CPU: 15 PID: 202299 Comm: mmap1 Not tainted 5.17.0-20220113.rc0.git0.f2211f194038.300.fc35.s390x+debug #1
Hardware name: IBM 3906 M04 704 (LPAR)
Call Trace:
[<00000001888aacfe>] dump_stack_lvl+0x76/0x98
[<0000000187c6d7be>] check_noncircular+0x136/0x158
[<0000000187c6e888>] check_prev_add+0xe0/0xed8
[<0000000187c6fdb6>] validate_chain+0x736/0xb20
[<0000000187c71e54>] __lock_acquire+0x604/0xbd8
[<0000000187c7301a>] lock_acquire.part.0+0xe2/0x238
[<0000000187c73220>] lock_acquire+0xb0/0x200
[<00000001888bf9aa>] _raw_spin_lock_irqsave+0x6a/0xd8
[<0000000187ef6862>] obj_cgroup_release+0x4a/0xe0
[<0000000187ef6498>] percpu_ref_put_many.constprop.0+0x150/0x168
[<0000000187ef9674>] drain_obj_stock+0x94/0xe8
[<0000000187efa464>] refill_obj_stock+0x94/0x278
[<0000000187eff55c>] obj_cgroup_charge+0x164/0x1d8
[<0000000187ed8aa4>] kmem_cache_alloc+0xac/0x528
[<0000000187bf2eb8>] __sigqueue_alloc+0x150/0x308
[<0000000187bf4210>] __send_signal+0x260/0x550
[<0000000187bf5f06>] send_signal+0x7e/0x348
[<0000000187bf7274>] force_sig_info_to_task+0x104/0x180
[<0000000187bf7758>] force_sig_fault+0x48/0x58
[<00000001888ae160>] __do_pgm_check+0x120/0x1f0
[<00000001888c0cde>] pgm_check_handler+0x11e/0x180
INFO: lockdep is turned off.

In this example a slab allocation from __send_signal() caused a
refilling and draining of a percpu objcg stock, resulted in a
releasing of another non-related objcg. Objcg release path requires
taking the css_set_lock, which is used to synchronize objcg lists.

This can create a circular dependency with the sighandler lock,
which is taken with the locked css_set_lock by the freezer code
(to freeze a task).

In general it seems that using css_set_lock to synchronize objcg lists
makes any slab allocations and deallocation with the locked
css_set_lock and any intervened locks risky.

To fix the problem and make the code more robust let's stop using
css_set_lock to synchronize objcg lists and use a new dedicated
spinlock instead.

Fixes: bf4f059954dc ("mm: memcg/slab: obj_cgroup API")
Signed-off-by: Roman Gushchin <guro@xxxxxx>
Reported-by: Alexander Egorenkov <egorenar@xxxxxxxxxxxxx>
Tested-by: Alexander Egorenkov <egorenar@xxxxxxxxxxxxx>
Reviewed-by: Waiman Long <longman@xxxxxxxxxx>
Cc: Tejun Heo <tj@xxxxxxxxxx>
Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
Cc: Shakeel Butt <shakeelb@xxxxxxxxxx>
Cc: Jeremy Linton <jeremy.linton@xxxxxxx>
Cc: cgroups@xxxxxxxxxxxxxxx
---
include/linux/memcontrol.h | 5 +++--
mm/memcontrol.c | 10 +++++-----
2 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index b72d75141e12..0abbd685703b 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -219,7 +219,7 @@ struct obj_cgroup {
struct mem_cgroup *memcg;
atomic_t nr_charged_bytes;
union {
- struct list_head list;
+ struct list_head list; /* protected by objcg_lock */
struct rcu_head rcu;
};
};
@@ -315,7 +315,8 @@ struct mem_cgroup {
#ifdef CONFIG_MEMCG_KMEM
int kmemcg_id;
struct obj_cgroup __rcu *objcg;
- struct list_head objcg_list; /* list of inherited objcgs */
+ /* list of inherited objcgs, protected by objcg_lock */
+ struct list_head objcg_list;
#endif
MEMCG_PADDING(_pad2_);
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 09d342c7cbd0..36e9f38c919d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -254,7 +254,7 @@ struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
}
#ifdef CONFIG_MEMCG_KMEM
-extern spinlock_t css_set_lock;
+static DEFINE_SPINLOCK(objcg_lock);
bool mem_cgroup_kmem_disabled(void)
{
@@ -298,9 +298,9 @@ static void obj_cgroup_release(struct percpu_ref *ref)
if (nr_pages)
obj_cgroup_uncharge_pages(objcg, nr_pages);
- spin_lock_irqsave(&css_set_lock, flags);
+ spin_lock_irqsave(&objcg_lock, flags);
list_del(&objcg->list);
- spin_unlock_irqrestore(&css_set_lock, flags);
+ spin_unlock_irqrestore(&objcg_lock, flags);
percpu_ref_exit(ref);
kfree_rcu(objcg, rcu);
@@ -332,7 +332,7 @@ static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
- spin_lock_irq(&css_set_lock);
+ spin_lock_irq(&objcg_lock);
/* 1) Ready to reparent active objcg. */
list_add(&objcg->list, &memcg->objcg_list);
@@ -342,7 +342,7 @@ static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
/* 3) Move already reparented objcgs to the parent's list */
list_splice(&memcg->objcg_list, &parent->objcg_list);
- spin_unlock_irq(&css_set_lock);
+ spin_unlock_irq(&objcg_lock);
percpu_ref_kill(&objcg->refcnt);
}


Thanks for taking care of this. Since it looks the same as my patch aside from the fact that I also defensivly converted the list_del to a list_del_rcu.


Reviewed-by: Jeremy Linton <jeremy.linton@xxxxxxx>

and

Tested-by: Jeremy Linton <jeremy.linton@xxxxxxx>


Thanks again,