From: Nanyong Sun sunnanyong@huawei.com
The commit d38a2b7a9c93 ("mm: memcg/slab: fix memory leak at non-root kmem_cache destroy") introduced a problem: If one thread destroy a kmem_cache A and another thread concurrently create a kmem_cache B, which is mergeable with A and has same size with A, the B may fail to create due to the duplicate sysfs node. The scenario in detail: 1) Thread 1 uses kmem_cache_destroy() to destroy kmem_cache A which is mergeable, it decreases A's refcount and if refcount is 0, then call memcg_set_kmem_cache_dying() which set A->memcg_params.dying = true, then unlock the slab_mutex and call flush_memcg_workqueue(), it may cost a while. Note: now the sysfs node(like '/kernel/slab/:0000248') of A is still present, it will be deleted in shutdown_cache() which will be called after flush_memcg_workqueue() is done and lock the slab_mutex again. 2) Now if thread 2 is coming, it use kmem_cache_create() to create B, which is mergeable with A(their size is same), it gain the lock of slab_mutex, then call __kmem_cache_alias() trying to find a mergeable node, because of the below added code in commit d38a2b7a9c93 ("mm: memcg/slab: fix memory leak at non-root kmem_cache destroy"), B is not mergeable with A whose memcg_params.dying is true.
int slab_unmergeable(struct kmem_cache *s) if (s->refcount < 0) return 1;
/* * Skip the dying kmem_cache. */ if (s->memcg_params.dying) return 1;
return 0; }
So B has to create its own sysfs node by calling: create_cache-> __kmem_cache_create-> sysfs_slab_add-> kobject_init_and_add Because B is mergeable itself, its filename of sysfs node is based on its size, like '/kernel/slab/:0000248', which is duplicate with A, and the sysfs node of A is still present now, so kobject_init_and_add() will return fail and result in kmem_cache_create() fail.
Concurrently modprobe and rmmod the two modules below can reproduce the issue quickly: nf_conntrack_expect, se_sess_cache. See call trace in the end.
LTS versions of v4.19.y and v5.4.y have this problem, whereas linux versions after v5.9 do not have this problem because the patchset: ("The new cgroup slab memory controller") almost refactored memcg slab.
A potential solution(this patch belongs): Just let the dying kmem_cache be mergeable, the slab_mutex lock can prevent the race between alias kmem_cache creating thread and root kmem_cache destroying thread. In the destroying thread, after flush_memcg_workqueue() is done, judge the refcount again, if someone reference it again during un-lock time, we don't need to destroy the kmem_cache completely, we can reuse it.
Another potential solution: revert the commit d38a2b7a9c93 ("mm: memcg/slab: fix memory leak at non-root kmem_cache destroy"), compare to the fail of kmem_cache_create, the memory leak in special scenario seems less harmful.
Call trace: sysfs: cannot create duplicate filename '/kernel/slab/:0000248' Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0 02/06/2015 Call trace: dump_backtrace+0x0/0x198 show_stack+0x24/0x30 dump_stack+0xb0/0x100 sysfs_warn_dup+0x6c/0x88 sysfs_create_dir_ns+0x104/0x120 kobject_add_internal+0xd0/0x378 kobject_init_and_add+0x90/0xd8 sysfs_slab_add+0x16c/0x2d0 __kmem_cache_create+0x16c/0x1d8 create_cache+0xbc/0x1f8 kmem_cache_create_usercopy+0x1a0/0x230 kmem_cache_create+0x50/0x68 init_se_kmem_caches+0x38/0x258 [target_core_mod] target_core_init_configfs+0x8c/0x390 [target_core_mod] do_one_initcall+0x54/0x230 do_init_module+0x64/0x1ec load_module+0x150c/0x16f0 __se_sys_finit_module+0xf0/0x108 __arm64_sys_finit_module+0x24/0x30 el0_svc_common+0x80/0x1c0 el0_svc_handler+0x78/0xe0 el0_svc+0x10/0x260 kobject_add_internal failed for :0000248 with -EEXIST, don't try to register things with the same name in the same directory. kmem_cache_create(se_sess_cache) failed with error -17 Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0 02/06/2015 Call trace: dump_backtrace+0x0/0x198 show_stack+0x24/0x30 dump_stack+0xb0/0x100 kmem_cache_create_usercopy+0xa8/0x230 kmem_cache_create+0x50/0x68 init_se_kmem_caches+0x38/0x258 [target_core_mod] target_core_init_configfs+0x8c/0x390 [target_core_mod] do_one_initcall+0x54/0x230 do_init_module+0x64/0x1ec load_module+0x150c/0x16f0 __se_sys_finit_module+0xf0/0x108 __arm64_sys_finit_module+0x24/0x30 el0_svc_common+0x80/0x1c0 el0_svc_handler+0x78/0xe0 el0_svc+0x10/0x260
Fixes: d38a2b7a9c93 ("mm: memcg/slab: fix memory leak at non-root kmem_cache destroy") Signed-off-by: Nanyong Sun sunnanyong@huawei.com Cc: stable@vger.kernel.org Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org --- mm/slab_common.c | 18 ++++++++++-------- 1 file changed, 10 insertions(+), 8 deletions(-)
--- a/mm/slab_common.c +++ b/mm/slab_common.c @@ -311,14 +311,6 @@ int slab_unmergeable(struct kmem_cache * if (s->refcount < 0) return 1;
-#ifdef CONFIG_MEMCG_KMEM - /* - * Skip the dying kmem_cache. - */ - if (s->memcg_params.dying) - return 1; -#endif - return 0; }
@@ -918,6 +910,16 @@ void kmem_cache_destroy(struct kmem_cach get_online_mems();
mutex_lock(&slab_mutex); + + /* + * Another thread referenced it again + */ + if (READ_ONCE(s->refcount)) { + spin_lock_irq(&memcg_kmem_wq_lock); + s->memcg_params.dying = false; + spin_unlock_irq(&memcg_kmem_wq_lock); + goto out_unlock; + } #endif
err = shutdown_memcg_caches(s);