[CCing Kirill] On Tue 21-06-16 22:05:56, zhongjiang wrote: > From: zhong jiang <[email protected]> > > with great pressure, I run some test cases. As a result, I found > that the THP is not freed, it is detected by check_mm(). > > BUG: Bad rss-counter state mm:ffff8827edb70000 idx:1 val:512 > > Consider the following race : > > CPU0 CPU1 > __handle_mm_fault() > wp_huge_pmd() > do_huge_pmd_wp_page() > pmdp_huge_clear_flush_notify() > (pmd_none = true) > exit_mmap() > unmap_vmas() > zap_pmd_range() > > pmd_none_or_trans_huge_or_clear_bad() > (result in memory leak) > set_pmd_at() > > because of CPU0 have allocated huge page before pmdp_huge_clear_notify, > and it make the pmd entry to be null. Therefore, The memory leak can occur.
I do not understand this description. CPU1 is in the exit path with last mm user gone. So CPU0 is a different process with its own mm. How can they influence each other. But maybe I am just missing your point. > The patch fix the scenario that the pmd entry can lead to be null. > > Signed-off-by: zhong jiang <[email protected]> > --- > mm/huge_memory.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c > index e10a4fe..ef04b94 100644 > --- a/mm/huge_memory.c > +++ b/mm/huge_memory.c > @@ -1340,11 +1340,11 @@ alloc: > pmd_t entry; > entry = mk_huge_pmd(new_page, vma->vm_page_prot); > entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); > - pmdp_huge_clear_flush_notify(vma, haddr, pmd); > + pmdp_invalidate(vma, haddr, pmd); > page_add_new_anon_rmap(new_page, vma, haddr, true); > mem_cgroup_commit_charge(new_page, memcg, false, true); > lru_cache_add_active_or_unevictable(new_page, vma); > - set_pmd_at(mm, haddr, pmd, entry); > + pmd_populate(mm, pmd, entry); > update_mmu_cache_pmd(vma, address, pmd); > if (!page) { > add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); > -- > 1.8.3.1 -- Michal Hocko SUSE Labs
