Skip to contentSkip to conversation

IB/ipoib fixes

PATCH Applied Daniel Vacek 2 participants 12 messages Dec 11, 2023 – Dec 17, 2023 InfiniBand

lore
Share & tools
lore
#0 lore raw

cover IB/ipoib fixes

The first patch (hopefully) fixes a real issue while the second is an unrelated cleanup. But it shares a context so sending as a series.

Daniel Vacek (2):
IB/ipoib: Fix mcast list locking
IB/ipoib: Clean up redundant netif_addr_lock

 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH 1/2 IB/ipoib: Fix mcast list locking

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(). 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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -0
index 5b3154503bf4..8e4f2c8839be 100644
@@ -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)
Signature · 1 line
2.43.0
#2 lore raw

Re: [PATCH 1/2] IB/ipoib: Fix mcast list locking

Quoted 28 lines · Daniel Vacek

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)
Quoted 1 line

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?

Thanks

Quoted 91 lines

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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -0
index 5b3154503bf4..8e4f2c8839be 100644
@@ -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)
Signature · 1 line
2.43.0
#3 lore raw

Re: [PATCH 1/2] IB/ipoib: Fix mcast list locking

Quoted 30 lines · Leon Romanovsky
Quoted 28 lines · Daniel Vacek

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)
Quoted 6 lines
Quoted 1 line

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.

--nX

Quoted 94 lines

Thanks

Quoted 91 lines

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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -1
index 5b3154503bf4..8e4f2c8839be 100644
@@ -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

#4 lore raw

Re: [PATCH 1/2] IB/ipoib: Fix mcast list locking

Quoted 44 lines · Daniel Vacek
Quoted 30 lines · Leon Romanovsky
Quoted 28 lines · Daniel Vacek

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)
Quoted 6 lines
Quoted 1 line

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.

Quoted 6 lines

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?

drivers/infiniband/ulp/ipoib/ipoib_multicast.c +1 -5
index 5b3154503bf4..bca80fe07584 100644
@@ -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;
 }
Quoted 99 lines

--nX

Quoted 94 lines

Thanks

Quoted 91 lines

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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -1
index 5b3154503bf4..8e4f2c8839be 100644
@@ -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

#5 lore raw

Re: [PATCH 1/2] IB/ipoib: Fix mcast list locking

Quoted 3 lines · Leon Romanovsky

What about the following change instead of adding extra lock to already too much complicated IPoIB?

Yeah, that's the other option should also work I believe. And it simplifies the code nicely.

The allocated mcast_member and mcast_group structures are small enough so that slab (by default) should not need more then order 1 block to eventually extend/refill the full kmalloc-256 cache. Some arches will even use order 0 I believe.
And unless I'm missing something I do not see any other sleeps in that path.

That said, as long as you are fine with occasional failures under memory pressure, it looks OK to me.

--nX

Quoted 129 lines
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +1 -5
index 5b3154503bf4..bca80fe07584 100644
@@ -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;
 }
Quoted 101 lines

--nX

Quoted 94 lines

Thanks

Quoted 91 lines

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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -1
index 5b3154503bf4..8e4f2c8839be 100644
@@ -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

#6 lore raw

Re: [PATCH 1/2] IB/ipoib: Fix mcast list locking

Quoted 8 lines · Daniel Vacek
Quoted 3 lines · Leon Romanovsky

What about the following change instead of adding extra lock to already too much complicated IPoIB?

Yeah, that's the other option should also work I believe. And it simplifies the code nicely.

Quoted 8 lines

The allocated mcast_member and mcast_group structures are small enough so that slab (by default) should not need more then order 1 block to eventually extend/refill the full kmalloc-256 cache. Some arches will even use order 0 I believe.
And unless I'm missing something I do not see any other sleeps in that path.

That said, as long as you are fine with occasional failures under memory pressure, it looks OK to me.

Yes, IMHO change from GFP_KERNEL to be GFP_ATOMIC is safer than adding extra lock.

Thanks

Quoted 133 lines

--nX

Quoted 129 lines
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +1 -5
index 5b3154503bf4..bca80fe07584 100644
@@ -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;
 }
Quoted 101 lines

--nX

Quoted 94 lines

Thanks

Quoted 91 lines

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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +3 -1
index 5b3154503bf4..8e4f2c8839be 100644
@@ -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

#7 lore raw

PATCH 2/2 IB/ipoib: Clean up redundant netif_addr_lock

In reply to Daniel Vacek

A single memory load does not need to be protected by any lock. The same priv->flags are fetched 15 lines ago without locking anyways.

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +1 -5
index 8e4f2c8839be..f54e0d212630 100644
@@ -572,13 +572,9 @@ void ipoib_mcast_join_task(struct work_struct *work)
 		return;
 	}
 	priv->local_lid = port_attr.lid;
-	netif_addr_lock_bh(dev);
 
-	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
-		netif_addr_unlock_bh(dev);
+	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags))
 		return;
-	}
-	netif_addr_unlock_bh(dev);
 
 	mutex_lock(&priv->mcast_mutex);
 	spin_lock_irq(&priv->lock);
Signature · 1 line
2.43.0
#8 lore raw

PATCH v2 1/1 IB/ipoib: Fix mcast list locking

In reply to Daniel Vacek

Releasing the priv->lock while iterating the priv->multicast_list in ipoib_mcast_join_task() opens a window for ipoib_mcast_dev_flush() to remove the items while in the middle of iteration. 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) |
        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   |  >>>  wait_for_completion(&mcast->done)
    the other thread which is blocked  |
    and the list is still valid on     |

it's stack.)

Fix this by keeping the lock held and changing to GFP_ATOMIC to prevent eventual sleeps.
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,

---
 drivers/infiniband/ulp/ipoib/ipoib_multicast.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/infiniband/ulp/ipoib/ipoib_multicast.c +1 -5
index 5b3154503bf4..bca80fe07584 100644
@@ -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;
 }
Signature · 1 line
2.43.0
#9 lore raw
In reply to Daniel Vacek
Quoted 8 lines · Daniel Vacek

The first patch (hopefully) fixes a real issue while the second is an unrelated cleanup. But it shares a context so sending as a series.

Daniel Vacek (2):
IB/ipoib: Fix mcast list locking
IB/ipoib: Clean up redundant netif_addr_lock

[...]

Applied, thanks!

[1/1] IB/ipoib: Fix mcast list locking

      https://git.kernel.org/rdma/rdma/c/4f973e211b3b1c

Best regards,

Signature · 1 line
Leon Romanovsky <[email protected]>
#10 lore raw
Quoted 6 lines · Leon Romanovsky
Quoted 3 lines · Daniel Vacek

The first patch (hopefully) fixes a real issue while the second is an unrelated cleanup. But it shares a context so sending as a series.

Quoted 7 lines
Quoted 5 lines

Daniel Vacek (2):
IB/ipoib: Fix mcast list locking
IB/ipoib: Clean up redundant netif_addr_lock

[...]

Applied, thanks!

Thank you.

One small detail - I was asked by Yuya to change the "Reported-by" as follows:

---

---

Would that be possible? And if yes, could you amend the commit yourself or do you want me to send a v3?

--nX

Quoted 7 lines

[1/1] IB/ipoib: Fix mcast list locking

      https://git.kernel.org/rdma/rdma/c/4f973e211b3b1c

Best regards,
--
Leon Romanovsky <[email protected]>

#11 lore raw
Quoted 17 lines · Daniel Vacek
Quoted 6 lines · Leon Romanovsky
Quoted 3 lines · Daniel Vacek

The first patch (hopefully) fixes a real issue while the second is an unrelated cleanup. But it shares a context so sending as a series.

Quoted 7 lines
Quoted 5 lines

Daniel Vacek (2):
IB/ipoib: Fix mcast list locking
IB/ipoib: Clean up redundant netif_addr_lock

[...]

Applied, thanks!

Thank you.

Quoted 8 lines

One small detail - I was asked by Yuya to change the "Reported-by" as follows:

---

---

Would that be possible? And if yes, could you amend the commit yourself or do you want me to send a v3?

Unfortunately, it is already too late as we promoted my wip branch to be official rdma/for-next.

Thanks

Quoted 12 lines

--nX

Quoted 7 lines

[1/1] IB/ipoib: Fix mcast list locking

      https://git.kernel.org/rdma/rdma/c/4f973e211b3b1c

Best regards,
--
Leon Romanovsky <[email protected]>