[PATCH v2 4.19 0/3] Backport patches to fix threadgroup_rwsem <-> cpus_read_lock() deadlock

Cai Xinchen posted 3 patches 2 years, 10 months ago
include/linux/cpuset.h    |  8 +++----
kernel/cgroup/cgroup-v1.c |  3 +++
kernel/cgroup/cgroup.c    | 50 +++++++++++++++++++++++++++++++++++----
kernel/cgroup/cpuset.c    | 25 ++++++++++++--------
4 files changed, 67 insertions(+), 19 deletions(-)
[PATCH v2 4.19 0/3] Backport patches to fix threadgroup_rwsem <-> cpus_read_lock() deadlock
Posted by Cai Xinchen 2 years, 10 months ago
I am very sorry. My gcc version is 7.5 and it does not report error.

We have a deadlock problem which can be solved by commit 4f7e7236435ca
("cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock").
However, it makes lock order of cpus_read_lock and cpuset_mutex
wrong in v4.19. The call sequence is as follows:
cgroup_procs_write()
        cgroup_procs_write_start()
                get_online_cpus(); // cpus_read_lock()
                percpu_down_write(&cgroup_threadgroup_rwsem)
        cgroup_attach_task
                cgroup_migrate
                        cgroup_migrate_execute
                                ss->attach (cpust_attach)
                                        mutex_lock(&cpuset_mutex)
it seems hard to make cpus_read_lock is locked before
cgroup_threadgroup_rwsem and cpuset_mutex is locked before
cpus_read_lock unless backport the commit d74b27d63a8beb
("cgroup/cpuset: Change cpuset_rwsem and hotplug lock order")

Changes in v2:
        * Add #include <linux/cpu.h> in kernel/cgroup/cgroup.c to
         avoid some compile error.
        * Exchange get_online_cpus() location in cpuset_attach to
         keep cpu_hotplug_lock->cpuset_mutex order, although it will
          be remove by ("cgroup: Fix threadgroup_rwsem <->
         cpus_read_lock() deadlock")

Juri Lelli (1):
  cgroup/cpuset: Change cpuset_rwsem and hotplug lock order

Tejun Heo (1):
  cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock

Tetsuo Handa (1):
  cgroup: Add missing cpus_read_lock() to cgroup_attach_task_all()

 include/linux/cpuset.h    |  8 +++----
 kernel/cgroup/cgroup-v1.c |  3 +++
 kernel/cgroup/cgroup.c    | 50 +++++++++++++++++++++++++++++++++++----
 kernel/cgroup/cpuset.c    | 25 ++++++++++++--------
 4 files changed, 67 insertions(+), 19 deletions(-)

-- 
2.17.1
Re: [PATCH v2 4.19 0/3] Backport patches to fix threadgroup_rwsem <-> cpus_read_lock() deadlock
Posted by Greg KH 2 years, 10 months ago
On Mon, Mar 20, 2023 at 01:15:04AM +0000, Cai Xinchen wrote:
> I am very sorry. My gcc version is 7.5 and it does not report error.
> 
> We have a deadlock problem which can be solved by commit 4f7e7236435ca
> ("cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock").
> However, it makes lock order of cpus_read_lock and cpuset_mutex
> wrong in v4.19. The call sequence is as follows:
> cgroup_procs_write()
>         cgroup_procs_write_start()
>                 get_online_cpus(); // cpus_read_lock()
>                 percpu_down_write(&cgroup_threadgroup_rwsem)
>         cgroup_attach_task
>                 cgroup_migrate
>                         cgroup_migrate_execute
>                                 ss->attach (cpust_attach)
>                                         mutex_lock(&cpuset_mutex)
> it seems hard to make cpus_read_lock is locked before
> cgroup_threadgroup_rwsem and cpuset_mutex is locked before
> cpus_read_lock unless backport the commit d74b27d63a8beb
> ("cgroup/cpuset: Change cpuset_rwsem and hotplug lock order")
> 
> Changes in v2:
>         * Add #include <linux/cpu.h> in kernel/cgroup/cgroup.c to
>          avoid some compile error.
>         * Exchange get_online_cpus() location in cpuset_attach to
>          keep cpu_hotplug_lock->cpuset_mutex order, although it will
>           be remove by ("cgroup: Fix threadgroup_rwsem <->
>          cpus_read_lock() deadlock")
> 
> Juri Lelli (1):
>   cgroup/cpuset: Change cpuset_rwsem and hotplug lock order
> 
> Tejun Heo (1):
>   cgroup: Fix threadgroup_rwsem <-> cpus_read_lock() deadlock
> 
> Tetsuo Handa (1):
>   cgroup: Add missing cpus_read_lock() to cgroup_attach_task_all()
> 
>  include/linux/cpuset.h    |  8 +++----
>  kernel/cgroup/cgroup-v1.c |  3 +++
>  kernel/cgroup/cgroup.c    | 50 +++++++++++++++++++++++++++++++++++----
>  kernel/cgroup/cpuset.c    | 25 ++++++++++++--------
>  4 files changed, 67 insertions(+), 19 deletions(-)

Now queued up, thanks.

greg k-h