summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorShakeel Butt <shakeel.butt@linux.dev>2024-08-02 16:58:22 -0700
committerAndrew Morton <akpm@linux-foundation.org>2024-08-07 18:33:56 -0700
commit9972605a238339b85bd16b084eed5f18414d22db (patch)
tree4d638d04802c3b067acad5ad2ec2e40bf6832a31
parent4cbf320b1500fe64fcef8c96ed74dfc1ae2c9e2c (diff)
memcg: protect concurrent access to mem_cgroup_idr
Commit 73f576c04b94 ("mm: memcontrol: fix cgroup creation failure after many small jobs") decoupled the memcg IDs from the CSS ID space to fix the cgroup creation failures. It introduced IDR to maintain the memcg ID space. The IDR depends on external synchronization mechanisms for modifications. For the mem_cgroup_idr, the idr_alloc() and idr_replace() happen within css callback and thus are protected through cgroup_mutex from concurrent modifications. However idr_remove() for mem_cgroup_idr was not protected against concurrency and can be run concurrently for different memcgs when they hit their refcnt to zero. Fix that. We have been seeing list_lru based kernel crashes at a low frequency in our fleet for a long time. These crashes were in different part of list_lru code including list_lru_add(), list_lru_del() and reparenting code. Upon further inspection, it looked like for a given object (dentry and inode), the super_block's list_lru didn't have list_lru_one for the memcg of that object. The initial suspicions were either the object is not allocated through kmem_cache_alloc_lru() or somehow memcg_list_lru_alloc() failed to allocate list_lru_one() for a memcg but returned success. No evidence were found for these cases. Looking more deeply, we started seeing situations where valid memcg's id is not present in mem_cgroup_idr and in some cases multiple valid memcgs have same id and mem_cgroup_idr is pointing to one of them. So, the most reasonable explanation is that these situations can happen due to race between multiple idr_remove() calls or race between idr_alloc()/idr_replace() and idr_remove(). These races are causing multiple memcgs to acquire the same ID and then offlining of one of them would cleanup list_lrus on the system for all of them. Later access from other memcgs to the list_lru cause crashes due to missing list_lru_one. Link: https://lkml.kernel.org/r/20240802235822.1830976-1-shakeel.butt@linux.dev Fixes: 73f576c04b94 ("mm: memcontrol: fix cgroup creation failure after many small jobs") Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev> Acked-by: Muchun Song <muchun.song@linux.dev> Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@suse.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
-rw-r--r--mm/memcontrol.c22
1 files changed, 20 insertions, 2 deletions
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 960371788687..f29157288b7d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -3386,11 +3386,28 @@ static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
#define MEM_CGROUP_ID_MAX ((1UL << MEM_CGROUP_ID_SHIFT) - 1)
static DEFINE_IDR(mem_cgroup_idr);
+static DEFINE_SPINLOCK(memcg_idr_lock);
+
+static int mem_cgroup_alloc_id(void)
+{
+ int ret;
+
+ idr_preload(GFP_KERNEL);
+ spin_lock(&memcg_idr_lock);
+ ret = idr_alloc(&mem_cgroup_idr, NULL, 1, MEM_CGROUP_ID_MAX + 1,
+ GFP_NOWAIT);
+ spin_unlock(&memcg_idr_lock);
+ idr_preload_end();
+ return ret;
+}
static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
{
if (memcg->id.id > 0) {
+ spin_lock(&memcg_idr_lock);
idr_remove(&mem_cgroup_idr, memcg->id.id);
+ spin_unlock(&memcg_idr_lock);
+
memcg->id.id = 0;
}
}
@@ -3524,8 +3541,7 @@ static struct mem_cgroup *mem_cgroup_alloc(struct mem_cgroup *parent)
if (!memcg)
return ERR_PTR(error);
- memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
- 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
+ memcg->id.id = mem_cgroup_alloc_id();
if (memcg->id.id < 0) {
error = memcg->id.id;
goto fail;
@@ -3667,7 +3683,9 @@ static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
* publish it here at the end of onlining. This matches the
* regular ID destruction during offlining.
*/
+ spin_lock(&memcg_idr_lock);
idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
+ spin_unlock(&memcg_idr_lock);
return 0;
offline_kmem: