Received-SPF: pass (google.com: domain of linux-kernel+bounces-196555-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Precedence: bulk
MIME-Version: 1.0
References: <20240531030520.1615833-1-zhaoyang.huang@unisoc.com> <ZlmEp9nxKiG9gWFj@pc636>
In-Reply-To: <ZlmEp9nxKiG9gWFj@pc636>
From: Zhaoyang Huang <huangzhaoyang@gmail.com>
Date: Fri, 31 May 2024 17:11:45 +0800
Message-ID: <CAGWkznGak0txoOEq1SYL9Ymax04Tac2nVCSYiC+L8qQ6bqryZQ@mail.gmail.com>
Subject: Re: [PATCHv3] mm: fix incorrect vbq reference in purge_fragmented_block
To: Uladzislau Rezki <urezki@gmail.com>
Cc: "zhaoyang.huang" <zhaoyang.huang@unisoc.com>, Andrew Morton <akpm@linux-foundation.org>, 
	Christoph Hellwig <hch@infradead.org>, Lorenzo Stoakes <lstoakes@gmail.com>, Baoquan He <bhe@redhat.com>, 
	Thomas Gleixner <tglx@linutronix.de>, hailong liu <hailong.liu@oppo.com>, linux-mm@kvack.org, 
	linux-kernel@vger.kernel.org, steve.kang@unisoc.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Fri, May 31, 2024 at 4:05=E2=80=AFPM Uladzislau Rezki <urezki@gmail.com>=
 wrote:
>
> On Fri, May 31, 2024 at 11:05:20AM +0800, zhaoyang.huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> > vmalloc area runs out in our ARM64 system during an erofs test as
> > vm_map_ram failed[1]. By following the debug log, we find that
> > vm_map_ram()->vb_alloc() will allocate new vb->va which corresponding
> > to 4MB vmalloc area as list_for_each_entry_rcu returns immediately
> > when vbq->free->next points to vbq->free. That is to say, 65536 times
> > of page fault after the list's broken will run out of the whole
> > vmalloc area. This should be introduced by one vbq->free->next point to
> > vbq->free which makes list_for_each_entry_rcu can not iterate the list
> > and find the BUG.
> >
> > [1]
> > PID: 1        TASK: ffffff80802b4e00  CPU: 6    COMMAND: "init"
> >  #0 [ffffffc08006afe0] __switch_to at ffffffc08111d5cc
> >  #1 [ffffffc08006b040] __schedule at ffffffc08111dde0
> >  #2 [ffffffc08006b0a0] schedule at ffffffc08111e294
> >  #3 [ffffffc08006b0d0] schedule_preempt_disabled at ffffffc08111e3f0
> >  #4 [ffffffc08006b140] __mutex_lock at ffffffc08112068c
> >  #5 [ffffffc08006b180] __mutex_lock_slowpath at ffffffc08111f8f8
> >  #6 [ffffffc08006b1a0] mutex_lock at ffffffc08111f834
> >  #7 [ffffffc08006b1d0] reclaim_and_purge_vmap_areas at ffffffc0803ebc3c
> >  #8 [ffffffc08006b290] alloc_vmap_area at ffffffc0803e83fc
> >  #9 [ffffffc08006b300] vm_map_ram at ffffffc0803e78c0
> >
> > Fixes: fc1e0d980037 ("mm/vmalloc: prevent stale TLBs in fully utilized =
blocks")
> >
> > Suggested-by: Hailong.Liu <hailong.liu@oppo.com>
> > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> Is a problem related to run out of vmalloc space _only_ or it is a proble=
m
> with broken list? From the commit message it is hard to follow the reason=
.
>
> Could you please post a full trace or panic?
Please refer to the below scenario for how vbq->free broken.
step 1: new_vmap_block is called in CPU0 and get vb->va->addr =3D
0xffffffc000400000
step 2: vb is added to CPU1's vbq->vmap_block(xarray) by xa =3D
addr_to_vb_xa(va->va_start);
            fc1e0d980037 ("mm/vmalloc: prevent stale TLBs in fully
utilized blocks") introduce a per_cpu like xarray mechanism to have vb
be added to the corresponding CPU's xarray but not local.
step 3: vb is added to CPU0's vbq->free by
list_add_tail_rcu(&vb->free_list, &vbq->free);
step 4 : purge_fragmented_blocks get vbq of CPU1 and then get above vb
step 5 : purge_fragmented_blocks delete vb from CPU0's list with
taking the vbq->lock of CPU1
step 5': vb_alloc on CPU0 could race with step5 and break the CPU0's vbq->f=
ree

As fc1e0d980037 solved the problem of staled TLB issue, we need to
introduce a new variable to record the CPU in vmap_block instead of
reverting to iterate the list(will leave wrong TLB entry)
>
> > ---
> > v2: introduce cpu in vmap_block to record the right CPU number
> > v3: use get_cpu/put_cpu to prevent schedule between core
> > ---
> > ---
> >  mm/vmalloc.c | 12 ++++++++----
> >  1 file changed, 8 insertions(+), 4 deletions(-)
> >
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 22aa63f4ef63..ecdb75d10949 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -2458,6 +2458,7 @@ struct vmap_block {
> >       struct list_head free_list;
> >       struct rcu_head rcu_head;
> >       struct list_head purge;
> > +     unsigned int cpu;
> >  };
> >
> >  /* Queue of free and dirty vmap blocks, for allocation and flushing pu=
rposes */
> > @@ -2586,10 +2587,12 @@ static void *new_vmap_block(unsigned int order,=
 gfp_t gfp_mask)
> >               return ERR_PTR(err);
> >       }
> >
> > +     vb->cpu =3D get_cpu();
> >       vbq =3D raw_cpu_ptr(&vmap_block_queue);
> >       spin_lock(&vbq->lock);
> >       list_add_tail_rcu(&vb->free_list, &vbq->free);
> >       spin_unlock(&vbq->lock);
> > +     put_cpu();
> >
> Why do you need get_cpu() here? Can you go with raw_smp_processor_id()
> and then access the per-cpu "vmap_block_queue"? get_cpu() disables
> preemption and then a spin-lock is take within this critical section.
> From the first glance PREEMPT_RT is broken in this case.
get_cpu here is to prevent current task from being migrated to other
COREs before we get the per_cpu vmap_block_queue. Could you please
suggest a correct way of doing this?

>
> I am on a vacation, responds can be with delays.
>
> --
> Uladzislau Rezki