From: Tang Yizhou tangyizhou@huawei.com

ascend inclusion category: bugfix bugzilla: NA CVE: NA

-------------------------------------------------

We found a concurrency problem of sp_group_add_task and sp_free which lead to memory leak.

After process A calls __sp_free and vfs_fallocate but before calling __sp_area_drop, process B is being added to the same group by a manager process, the *dead* spa freed by sp_free may be mapped into process B again, then do_mm_populate is called.

When sp_group_add_task is finished, this spa is dropped and can't be seen in /proc/sharepool/spa_stat, but the memory of spa still reside in the group. It can only be freed when the group is dead.

To fix the problem, we add a member is_dead in spa. We can access it when spg->rw_lock is held. This may sound a little strange if not realizing the life cycle of spa has a direct relation with sp group.

Suggested-by: Ding Tianhong dingtianhong@huawei.com Signed-off-by: Tang Yizhou tangyizhou@huawei.com Reviewed-by: Ding Tianhong dingtianhong@huwei.com Reviewed-by: 为珑 陈 chenweilong@huawei.com Signed-off-by: Yang Yingliang yangyingliang@huawei.com --- mm/share_pool.c | 37 +++++++++++++++++++++++++------------ 1 file changed, 25 insertions(+), 12 deletions(-)

diff --git a/mm/share_pool.c b/mm/share_pool.c index 61bbbd772c847..3e4e05f20c5ce 100644 --- a/mm/share_pool.c +++ b/mm/share_pool.c @@ -233,6 +233,7 @@ struct sp_area { unsigned long region_vstart; /* belong to normal region or DVPP region */ unsigned long flags; bool is_hugepage; + bool is_dead; atomic_t use_count; /* How many vmas use this VA region */ struct rb_node rb_node; /* address sorted rbtree */ struct list_head link; /* link to the spg->head */ @@ -736,6 +737,10 @@ int sp_group_add_task(int pid, int spg_id) prev = spa;

atomic_inc(&spa->use_count); + + if (spa->is_dead == true) + continue; + spin_unlock(&sp_area_lock);

if (spa->type == SPA_TYPE_K2SPG && spa->kva) { @@ -970,6 +975,7 @@ static struct sp_area *sp_alloc_area(unsigned long size, unsigned long flags, spa->region_vstart = vstart; spa->flags = flags; spa->is_hugepage = (flags & SP_HUGEPAGE); + spa->is_dead = false; spa->spg = spg; atomic_set(&spa->use_count, 1); spa->type = type; @@ -1271,10 +1277,14 @@ int sp_free(unsigned long addr) goto drop_spa; }

- if (!spg_valid(spa->spg)) + down_write(&spa->spg->rw_lock); + if (!spg_valid(spa->spg)) { + up_write(&spa->spg->rw_lock); goto drop_spa; - - sp_dump_stack(); + } + /* the life cycle of spa has a direct relation with sp group */ + spa->is_dead = true; + up_write(&spa->spg->rw_lock);

down_read(&spa->spg->rw_lock);

@@ -1303,6 +1313,7 @@ int sp_free(unsigned long addr) drop_spa: __sp_area_drop(spa); out: + sp_dump_stack(); sp_try_to_compact(); return ret; } @@ -2362,15 +2373,6 @@ static int sp_unshare_uva(unsigned long uva, unsigned long size, int pid, int sp goto out_drop_area; }

- down_read(&spa->spg->rw_lock); - if (!spg_valid(spa->spg)) { - up_read(&spa->spg->rw_lock); - pr_info_ratelimited("share pool: no need to unshare uva(to group), " - "sp group of spa is dead\n"); - goto out_clr_flag; - } - up_read(&spa->spg->rw_lock); - /* alway allow kthread and dvpp channel destroy procedure */ if (current->mm && current->mm->sp_group != spa->spg) { pr_err_ratelimited("share pool: unshare uva(to group) failed, " @@ -2379,6 +2381,17 @@ static int sp_unshare_uva(unsigned long uva, unsigned long size, int pid, int sp goto out_drop_area; }

+ down_write(&spa->spg->rw_lock); + if (!spg_valid(spa->spg)) { + up_write(&spa->spg->rw_lock); + pr_info_ratelimited("share pool: no need to unshare uva(to group), " + "sp group of spa is dead\n"); + goto out_clr_flag; + } + /* the life cycle of spa has a direct relation with sp group */ + spa->is_dead = true; + up_write(&spa->spg->rw_lock); + down_read(&spa->spg->rw_lock); __sp_free(spa->spg, uva_aligned, size_aligned, NULL); up_read(&spa->spg->rw_lock);