| Message ID | 20220802013251.3022141-1-mawupeng1@huawei.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [stable,4.14,4.19,1/1] mm: Fix page counter mismatch in shmem_mfill_atomic_pte | expand |
Cc Greg On 2022/8/2 9:32, Wupeng Ma wrote: > From: Ma Wupeng <mawupeng1@huawei.com> > > shmem_mfill_atomic_pte() wrongly called mem_cgroup_cancel_charge() in "success" > path, it should mem_cgroup_uncharge() to dec memory counter instead. > mem_cgroup_cancel_charge() should only be used if this transaction is > unsuccessful and mem_cgroup_uncharge() is used to do this if this transaction > succeed. > > This will lead to page->memcg not null and will uncharge one more in put_page(). > The page counter will underflow to maximum value and trigger oom to kill all > process include sshd and leave system unaccessible. > > page->memcg is set in the following path: > mem_cgroup_commit_charge > commit_charge > page->mem_cgroup = memcg; > > extra uncharge will be done in the following path: > put_page > __put_page > __put_single_page > mem_cgroup_uncharge > if (!page->mem_cgroup) <-- should return here > return > uncharge_page > uncharge_batch > > To fix this, call mem_cgroup_commit_charge() at the end of this transaction to > make sure this transaction is really finished. > > Fixes: 4c27fe4c4c84 ("userfaultfd: shmem: add shmem_mcopy_atomic_pte for userfaultfd support") > Signed-off-by: Ma Wupeng <mawupeng1@huawei.com> > --- > mm/shmem.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/mm/shmem.c b/mm/shmem.c > index 0788616696dc..0b06724c189e 100644 > --- a/mm/shmem.c > +++ b/mm/shmem.c > @@ -2339,8 +2339,6 @@ static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm, > if (ret) > goto out_release_uncharge; > > - mem_cgroup_commit_charge(page, memcg, false, false); > - > _dst_pte = mk_pte(page, dst_vma->vm_page_prot); > if (dst_vma->vm_flags & VM_WRITE) > _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte)); > @@ -2366,6 +2364,8 @@ static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm, > if (!pte_none(*dst_pte)) > goto out_release_uncharge_unlock; > > + mem_cgroup_commit_charge(page, memcg, false, false); > + > lru_cache_add_anon(page); > > spin_lock_irq(&info->lock);
On Tue, Aug 16, 2022 at 11:27:08AM +0800, mawupeng wrote:
> Cc Greg
Cc Greg for what? I have no context here at all as to what you want me
to do...
totally confused,
greg k-h
On 2022/8/16 13:31, Greg KH wrote: > On Tue, Aug 16, 2022 at 11:27:08AM +0800, mawupeng wrote: >> Cc Greg > > Cc Greg for what? I have no context here at all as to what you want me > to do.. We found a bug related to memory cgroup counter in stable 4.14/4.19. shmem_mfill_atomic_pte() wrongly called mem_cgroup_cancel_charge() in "success" path, it should mem_cgroup_uncharge() to dec memory counter instead. mem_cgroup_cancel_charge() should only be used if this transaction is unsuccessful and mem_cgroup_uncharge() is used to do this if this transaction succeed. Commit 3fea5a499d57 ("mm: memcontrol: convert page cache to a new mem_cgroup_charge() API") in v5.8-rc1 change is charge/uncharge/cancel logic so don't have this problem. This counter will underflow to negative maximum value and trigger oom to kill all process include sshd and leave system unaccessible. The reason cc you is that we want to merge this bugfix into stable 4.14/4.19. The error call trace: ------------[ cut here ]------------ WARNING: CPU: 0 PID: 17127 at mm/page_counter.c:62 page_counter_cancel+0x57/0x90 RIP: 0010:page_counter_cancel+0x57/0x90 Call Trace: page_counter_uncharge+0x33/0x60 uncharge_batch+0xb5/0x5f0 mem_cgroup_uncharge_list+0x102/0x170 release_pages+0x814/0xcc0 tlb_flush_mmu_free+0xa9/0x140 arch_tlb_finish_mmu+0xa4/0x140 tlb_finish_mmu+0x90/0xf0 exit_mmap+0x264/0x4b0 > > totally confused, > > greg k-h
On Tue, Aug 16, 2022 at 03:04:08PM +0800, mawupeng wrote: > > > On 2022/8/16 13:31, Greg KH wrote: > > On Tue, Aug 16, 2022 at 11:27:08AM +0800, mawupeng wrote: > >> Cc Greg > > > > Cc Greg for what? I have no context here at all as to what you want me > > to do.. > > We found a bug related to memory cgroup counter in stable 4.14/4.19. > shmem_mfill_atomic_pte() wrongly called mem_cgroup_cancel_charge() in "success" > path, it should mem_cgroup_uncharge() to dec memory counter instead. > mem_cgroup_cancel_charge() should only be used if this transaction is > unsuccessful and mem_cgroup_uncharge() is used to do this if this transaction > succeed. > > Commit 3fea5a499d57 ("mm: memcontrol: convert page cache to a new mem_cgroup_charge() API") > in v5.8-rc1 change is charge/uncharge/cancel logic so don't have this > problem. > > This counter will underflow to negative maximum value and trigger oom to kill all > process include sshd and leave system unaccessible. > > The reason cc you is that we want to merge this bugfix into stable 4.14/4.19. <formletter> This is not the correct way to submit patches for inclusion in the stable kernel tree. Please read: https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html for how to do this properly. </formletter>
diff --git a/mm/shmem.c b/mm/shmem.c index 0788616696dc..0b06724c189e 100644 --- a/mm/shmem.c +++ b/mm/shmem.c @@ -2339,8 +2339,6 @@ static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm, if (ret) goto out_release_uncharge; - mem_cgroup_commit_charge(page, memcg, false, false); - _dst_pte = mk_pte(page, dst_vma->vm_page_prot); if (dst_vma->vm_flags & VM_WRITE) _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte)); @@ -2366,6 +2364,8 @@ static int shmem_mfill_atomic_pte(struct mm_struct *dst_mm, if (!pte_none(*dst_pte)) goto out_release_uncharge_unlock; + mem_cgroup_commit_charge(page, memcg, false, false); + lru_cache_add_anon(page); spin_lock_irq(&info->lock);