Re: [PATCH 1/2] IB/ipoib: Fix mcast list locking
From: Leon Romanovsky
Date: Mon Dec 11 2023 - 10:07:06 EST
On Mon, Dec 11, 2023 at 03:25:39PM +0100, Daniel Vacek wrote:
> On Mon, Dec 11, 2023 at 2:45 PM Leon Romanovsky <leon@xxxxxxxxxx> wrote:
> >
> > On Mon, Dec 11, 2023 at 02:04:24PM +0100, Daniel Vacek wrote:
> > > We need an additional protection against list removal between ipoib_mcast_join_task()
> > > and ipoib_mcast_dev_flush() in case the &priv->lock needs to be dropped while
> > > iterating the &priv->multicast_list in ipoib_mcast_join_task(). If the mcast
> > > is removed while the lock was dropped, the for loop spins forever resulting
> > > in a hard lockup (as was reported on RHEL 4.18.0-372.75.1.el8_6 kernel):
> > >
> > > Task A (kworker/u72:2 below) | Task B (kworker/u72:0 below)
> > > -----------------------------------+-----------------------------------
> > > ipoib_mcast_join_task(work) | ipoib_ib_dev_flush_light(work)
> > > spin_lock_irq(&priv->lock) | __ipoib_ib_dev_flush(priv, ...)
> > > list_for_each_entry(mcast, | ipoib_mcast_dev_flush(dev = priv->dev)
> > > &priv->multicast_list, list) | mutex_lock(&priv->mcast_mutex)
> > > ipoib_mcast_join(dev, mcast) |
> > > spin_unlock_irq(&priv->lock) |
> > > | spin_lock_irqsave(&priv->lock, flags)
> > > | list_for_each_entry_safe(mcast, tmcast,
> > > | &priv->multicast_list, list)
> > > | list_del(&mcast->list);
> > > | list_add_tail(&mcast->list, &remove_list)
> > > | spin_unlock_irqrestore(&priv->lock, flags)
> > > spin_lock_irq(&priv->lock) |
> > > | ipoib_mcast_remove_list(&remove_list)
> > > (Here, mcast is no longer on the | list_for_each_entry_safe(mcast, tmcast,
> > > &priv->multicast_list and we keep | remove_list, list)
> > > spinning on the &remove_list of the \ >>> wait_for_completion(&mcast->done)
> > > other thread which is blocked and the|
> > > list is still valid on it's stack.) | mutex_unlock(&priv->mcast_mutex)
> > >
> > > Fix this by adding mutex_lock(&priv->mcast_mutex) to ipoib_mcast_join_task().
> >
> > I don't entirely understand the issue and the proposed solution.
> > There is only one spin_unlock_irq() in the middle of list_for_each_entry(mcast, &priv->multicast_list, list)
> > and it is right before return statement which will break the loop. So
> > how will loop spin forever?
>
> There's another unlock/lock pair around ib_sa_join_multicast() call in
> ipoib_mcast_join() no matter the outcome of the condition. The
> ib_sa_join_multicast() cannot be called with the lock being held due
> to GFP_KERNEL allocation can possibly sleep. That's what's causing the
> issue.
>
> Actually if you check the code, only if the mentioned condition is
> false (and the loop is not broken and returned from
> ipoib_mcast_join_task()) the lock is released and re-acquired,
> creating the window for
> ipoib_ib_dev_flush_light()/ipoib_mcast_dev_flush() to break the list.
> The vmcore data shows/confirms that clearly.
Thanks, it is more clear now.
What about the following change instead of adding extra lock to already
too much complicated IPoIB?
diff --git a/drivers/infiniband/ulp/ipoib/ipoib_multicast.c b/drivers/infiniband/ulp/ipoib/ipoib_multicast.c
index 5b3154503bf4..bca80fe07584 100644
--- a/drivers/infiniband/ulp/ipoib/ipoib_multicast.c
+++ b/drivers/infiniband/ulp/ipoib/ipoib_multicast.c
@@ -531,21 +531,17 @@ static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
rec.join_state = SENDONLY_FULLMEMBER_JOIN;
}
- spin_unlock_irq(&priv->lock);
multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
- &rec, comp_mask, GFP_KERNEL,
+ &rec, comp_mask, GFP_ATOMIC,
ipoib_mcast_join_complete, mcast);
- spin_lock_irq(&priv->lock);
if (IS_ERR(multicast)) {
ret = PTR_ERR(multicast);
ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
/* Requeue this join task with a backoff delay */
__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
- spin_unlock_irq(&priv->lock);
complete(&mcast->done);
- spin_lock_irq(&priv->lock);
}
return 0;
}
>
> --nX
>
>
> > Thanks
> >
> > > Unfortunately we could not reproduce the lockup and confirm this fix but
> > > based on the code review I think this fix should address such lockups.
> > >
> > > crash> bc 31
> > > PID: 747 TASK: ff1c6a1a007e8000 CPU: 31 COMMAND: "kworker/u72:2"
> > > --
> > > [exception RIP: ipoib_mcast_join_task+0x1b1]
> > > RIP: ffffffffc0944ac1 RSP: ff646f199a8c7e00 RFLAGS: 00000002
> > > RAX: 0000000000000000 RBX: ff1c6a1a04dc82f8 RCX: 0000000000000000
> > > work (&priv->mcast_task{,.work})
> > > RDX: ff1c6a192d60ac68 RSI: 0000000000000286 RDI: ff1c6a1a04dc8000
> > > &mcast->list
> > > RBP: ff646f199a8c7e90 R8: ff1c699980019420 R9: ff1c6a1920c9a000
> > > R10: ff646f199a8c7e00 R11: ff1c6a191a7d9800 R12: ff1c6a192d60ac00
> > > mcast
> > > R13: ff1c6a1d82200000 R14: ff1c6a1a04dc8000 R15: ff1c6a1a04dc82d8
> > > dev priv (&priv->lock) &priv->multicast_list (aka head)
> > > ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
> > > --- <NMI exception stack> ---
> > > #5 [ff646f199a8c7e00] ipoib_mcast_join_task+0x1b1 at ffffffffc0944ac1 [ib_ipoib]
> > > #6 [ff646f199a8c7e98] process_one_work+0x1a7 at ffffffff9bf10967
> > >
> > > crash> rx ff646f199a8c7e68
> > > ff646f199a8c7e68: ff1c6a1a04dc82f8 <<< work = &priv->mcast_task.work
> > >
> > > crash> list -hO ipoib_dev_priv.multicast_list ff1c6a1a04dc8000
> > > (empty)
> > >
> > > crash> ipoib_dev_priv.mcast_task.work.func,mcast_mutex.owner.counter ff1c6a1a04dc8000
> > > mcast_task.work.func = 0xffffffffc0944910 <ipoib_mcast_join_task>,
> > > mcast_mutex.owner.counter = 0xff1c69998efec000
> > >
> > > crash> b 8
> > > PID: 8 TASK: ff1c69998efec000 CPU: 33 COMMAND: "kworker/u72:0"
> > > --
> > > #3 [ff646f1980153d50] wait_for_completion+0x96 at ffffffff9c7d7646
> > > #4 [ff646f1980153d90] ipoib_mcast_remove_list+0x56 at ffffffffc0944dc6 [ib_ipoib]
> > > #5 [ff646f1980153de8] ipoib_mcast_dev_flush+0x1a7 at ffffffffc09455a7 [ib_ipoib]
> > > #6 [ff646f1980153e58] __ipoib_ib_dev_flush+0x1a4 at ffffffffc09431a4 [ib_ipoib]
> > > #7 [ff646f1980153e98] process_one_work+0x1a7 at ffffffff9bf10967
> > >
> > > crash> rx ff646f1980153e68
> > > ff646f1980153e68: ff1c6a1a04dc83f0 <<< work = &priv->flush_light
> > >
> > > crash> ipoib_dev_priv.flush_light.func,broadcast ff1c6a1a04dc8000
> > > flush_light.func = 0xffffffffc0943820 <ipoib_ib_dev_flush_light>,
> > > broadcast = 0x0,
> > >
> > > The mcast(s) on the &remove_list (the remaining part of the ex &priv->multicast_list):
> > >
> > > crash> list -s ipoib_mcast.done.done ipoib_mcast.list -H ff646f1980153e10 | paste - -
> > > ff1c6a192bd0c200 done.done = 0x0,
> > > ff1c6a192d60ac00 done.done = 0x0,
> > >
> > > Reported-by: Yuya Fujita-bishamonten <fj-lsoft-rh-driver@xxxxxxxxxxxxxxxxx>
> > > Signed-off-by: Daniel Vacek <neelx@xxxxxxxxxx>
> > > ---
> > > drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
> > > 1 file changed, 3 insertions(+)
> > >
> > > diff --git a/drivers/infiniband/ulp/ipoib/ipoib_multicast.c b/drivers/infiniband/ulp/ipoib/ipoib_multicast.c
> > > index 5b3154503bf4..8e4f2c8839be 100644
> > > --- a/drivers/infiniband/ulp/ipoib/ipoib_multicast.c
> > > +++ b/drivers/infiniband/ulp/ipoib/ipoib_multicast.c
> > > @@ -580,6 +580,7 @@ void ipoib_mcast_join_task(struct work_struct *work)
> > > }
> > > netif_addr_unlock_bh(dev);
> > >
> > > + mutex_lock(&priv->mcast_mutex);
> > > spin_lock_irq(&priv->lock);
> > > if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
> > > goto out;
> > > @@ -634,6 +635,7 @@ void ipoib_mcast_join_task(struct work_struct *work)
> > > /* Found the next unjoined group */
> > > if (ipoib_mcast_join(dev, mcast)) {
> > > spin_unlock_irq(&priv->lock);
> > > + mutex_unlock(&priv->mcast_mutex);
> > > return;
> > > }
> > > } else if (!delay_until ||
> > > @@ -655,6 +657,7 @@ void ipoib_mcast_join_task(struct work_struct *work)
> > > ipoib_mcast_join(dev, mcast);
> > >
> > > spin_unlock_irq(&priv->lock);
> > > + mutex_unlock(&priv->mcast_mutex);
> > > }
> > >
> > > void ipoib_mcast_start_thread(struct net_device *dev)
> > > --
> > > 2.43.0
> > >
> >
>