MLX4 Cq Question

I'm looking at the Linux MLX4 net driver and found something that confuses me mightily. In particular in the file net/ethernet/mellanox/mlx4/cq.c, the mlx4_ib_completion function does not take any kind of lock when looking up the SW CQ in the radix tree, however, the mlx4_cq_event function does. In addition if I go look at the code paths where cq are removed from this tree, they are protected by spin_lock_irq. So I am baffled at this point as to what the locking strategy is and how this is supposed to work. I'm sure I'm missing something and would greatly appreciate it if someone would explain this.

you see that when freeing a CQ, we first do the HW2SW_CQ firmware command; once this command completes, no more events will be generated for that CQ. Then we do synchronize_irq for the CQ's interrupt vector. Once that completes, no more completion handlers will be running for the CQ, so we can safely delete the CQ from the radix tree (relying on the radix tree's safety of deleting one entry while possibly looking up other entries, so no lock is needed). We also use the lock to synchronize against the CQ event function, which as you noted does take the lock too.

Basic idea is that we're tricky and careful so we can make the fast path (completion interrupt handling) lock-free, but then use locks and whatever else needed in the slow path (CQ async event handling, CQ destroy).

Quoted 1 line · Tom Tucker

I'm looking at the Linux MLX4 net driver and found something that confuses me mightily. In particular in the file net/ethernet/mellanox/mlx4/cq.c, the mlx4_ib_completion function does not take any kind of lock when looking up the SW CQ in the radix tree, however, the mlx4_cq_event function does. In addition if I go look at the code paths where cq are removed from this tree, they are protected by spin_lock_irq. So I am baffled at this point as to what the locking strategy is and how this is supposed to work. I'm sure I'm missing something and would greatly appreciate it if someone would explain this.

This is a bit tricky. If you look at

void mlx4_cq_free(struct mlx4_dev *dev, struct mlx4_cq *cq)

{
        struct mlx4_priv *priv = mlx4_priv(dev);
        struct mlx4_cq_table *cq_table = &priv->cq_table;
        int err;

        err = mlx4_HW2SW_CQ(dev, NULL, cq->cqn);
        if (err)
                mlx4_warn(dev, "HW2SW_CQ failed (%d) for CQN %06x\n",

err, cq->cqn);

        synchronize_irq(priv->eq_table.eq[cq->vector].irq);

        spin_lock_irq(&cq_table->lock);
        radix_tree_delete(&cq_table->tree, cq->cqn);
        spin_unlock_irq(&cq_table->lock);

        if (atomic_dec_and_test(&cq->refcount))
                complete(&cq->free);
        wait_for_completion(&cq->free);

        mlx4_cq_free_icm(dev, cq->cqn);
}

you see that when freeing a CQ, we first do the HW2SW_CQ firmware command; once this command completes, no more events will be generated for that CQ. Then we do synchronize_irq for the CQ's interrupt vector. Once that completes, no more completion handlers will be running for the CQ, so we can safely delete the CQ from the radix tree (relying on the radix tree's safety of deleting one entry while possibly looking up other entries, so no lock is needed). We also use the lock to synchronize against the CQ event function, which as you noted does take the lock too.

Basic idea is that we're tricky and careful so we can make the fast path (completion interrupt handling) lock-free, but then use locks and whatever else needed in the slow path (CQ async event handling, CQ destroy).

- R.

This is racy and can cause use-after-free, null pointer dereference, etc, which result in kernel crashes.

Quoted 38 lines · Roland Dreier

Quoted 1 line · Tom Tucker

I'm looking at the Linux MLX4 net driver and found something that confuses me mightily. In particular in the file net/ethernet/mellanox/mlx4/cq.c, the mlx4_ib_completion function does not take any kind of lock when looking up the SW CQ in the radix tree, however, the mlx4_cq_event function does. In addition if I go look at the code paths where cq are removed from this tree, they are protected by spin_lock_irq. So I am baffled at this point as to what the locking strategy is and how this is supposed to work. I'm sure I'm missing something and would greatly appreciate it if someone would explain this.

This is a bit tricky. If you look at

void mlx4_cq_free(struct mlx4_dev *dev, struct mlx4_cq *cq)

{
         struct mlx4_priv *priv = mlx4_priv(dev);
         struct mlx4_cq_table *cq_table = &priv->cq_table;
         int err;

         err = mlx4_HW2SW_CQ(dev, NULL, cq->cqn);
         if (err)
                 mlx4_warn(dev, "HW2SW_CQ failed (%d) for CQN %06x\n",

err, cq->cqn);

         synchronize_irq(priv->eq_table.eq[cq->vector].irq);

         spin_lock_irq(&cq_table->lock);
         radix_tree_delete(&cq_table->tree, cq->cqn);
         spin_unlock_irq(&cq_table->lock);

         if (atomic_dec_and_test(&cq->refcount))
                 complete(&cq->free);
         wait_for_completion(&cq->free);

         mlx4_cq_free_icm(dev, cq->cqn);
}

you see that when freeing a CQ, we first do the HW2SW_CQ firmware command; once this command completes, no more events will be generated for that CQ. Then we do synchronize_irq for the CQ's interrupt vector. Once that completes, no more completion handlers will be running for the CQ, so we can safely delete the CQ from the radix tree (relying on the radix tree's safety of deleting one entry while possibly looking up other entries, so no lock is needed). We also use the lock to synchronize against the CQ event function, which as you noted does take the lock too.

Quoted 6 lines

Basic idea is that we're tricky and careful so we can make the fast path (completion interrupt handling) lock-free, but then use locks and whatever else needed in the slow path (CQ async event handling, CQ destroy).

- R.

=======================================================

Roland, unfortunately we have seen that we need some locking on the cq completion handler (there is a stack trace which resulted from this lack of proper locking).
In our current driver, we are using the patch below (which uses RCU locking instead of spinlocks). I can prepare a proper patch for the upstream kernel.

===================================================
net/mlx4_core: Fix racy flow in the driver CQ completion handler

The mlx4 CQ completion handler, mlx4_cq_completion, doesn't bother to lock the radix tree which is used to manage the table of CQs, nor does it increase the reference count of the CQ before invoking the user provided callback (and decrease it afterwards).

This is racy and can cause use-after-free, null pointer dereference, etc, which result in kernel crashes.

To fix this, we must do the following in mlx4_cq_completion:
- increase the ref count on the cq before invoking the user callback, and decrement it after the callback.
- Place a lock around the radix tree lookup/ref-count-increase

Using an irq spinlock will not fix this issue. The problem is that under VPI, the ETH interface uses multiple msix irq's, which can result in one cq completion event interrupting another in-progress cq completion event. A deadlock results when the handler for the first cq completion grabs the spinlock, and is interrupted by the second completion before it has a chance to release the spinlock. The handler for the second completion will deadlock waiting for the spinlock to be released.

The proper fix is to use the RCU mechanism for locking radix-tree accesses in the cq completion event handler (The radix-tree implementation uses the RCU mechanism, so rcu_read_lock/unlock in the reader, with rcu_synchronize in the updater, will do the job).

Note that the same issue exists in mlx4_cq_event() (the cq async event handler), which also takes the same lock on the radix tree. Here, we replace the spinlock with an rcu_read_lock().

This patch was motivated by the following report from the field:

"[...] box panic'ed when trying to find a completion queue. There is no corruption but there is a possible race which could result in mlx4_cq_completion getting wrong height of the radix tree and following a bit too deep into the chains. In the other code which uses this radix tree the access is protected by the lock but mlx4_cq_completion is running in the interrupt context and cannot take locks, so instead it runs without any protection whatsoever."

The stack trace below is from the mlnx ofed 1.5.3 driver running under RHEL5.7. (this driver uses the upstream kernel mlx4_cq_completion() code)

PID: 8178 TASK: ffff810b91a52400 CPU: 11 COMMAND: "universal_consu"

[exception RIP: radix_tree_lookup+38]

RIP: ffffffff8016405f  RSP: ffff81182fe3be90  RFLAGS: 00210002
RAX: 6b6b6b6b6b6b6b8b  RBX: ffff810c2fb90000  RCX: 0000000000000006
RDX: 0000000000000001  RSI: 00000000000000c0  RDI: 6b6b6b6b6b6b6b6b
RBP: 00000000000000c0   R8: 0000000000010000   R9: 000000000920ea94
R10: 00000000000000b3  R11: ffffffff800c7ea5  R12: 00000000000000b3
R13: ffff810c2b15a280  R14: ffff810b7a98ff58  R15: 0000000000000000
ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0000

*** <IRQ stack> ***

RIP: 00000000f5cd8859  RSP: 00000000f3eafa28  RFLAGS: 00200202
RAX: 0000000000000000  RBX: 00000000f3eafaf8  RCX: 000000000b6d4f00
RDX: 000000000000001c  RSI: 000000000000001c  RDI: 0000000000000000
RBP: ffff810bf1a0a120   R8: 0000000000000000   R9: 0000000000000000
R10: 0000000000000000  R11: 0000000000000000  R12: 0000000000000001
R13: 0000000000000000  R14: ffff810b9e7d5bf0  R15: 0000000000000000
ORIG_RAX: ffffffffffffff4c  CS: 0023  SS: 002b

The original patch was generated by Yishai Hadas,
and reviewed by Or Gerlitz and Jack Morgenstein. A subsequent fix by Jack Morgenstein was reviewed by Eli Cohen.

Signed-off-by Jack Morgenstein <[email protected]>

---
  drivers/net/ethernet/mellanox/mlx4/cq.c |   14 ++++++++++++--
  1 files changed, 12 insertions(+), 2 deletions(-)

drivers/net/ethernet/mellanox/mlx4/cq.c +11 -2

index ff91904..0e28258 100644
@@ -57,8 +57,13 @@ void mlx4_cq_completion(struct mlx4_dev *dev, u32 cqn)
  {
  	struct mlx4_cq *cq;
 
+	rcu_read_lock();
  	cq = radix_tree_lookup(&mlx4_priv(dev)->cq_table.tree,
  			       cqn & (dev->caps.num_cqs - 1));
+	if (cq)
+		atomic_inc(&cq->refcount);
+	rcu_read_unlock();
+
  	if (!cq) {
  		mlx4_dbg(dev, "Completion event for bogus CQ %08x\n", cqn);
  		return;

@@ -67,6 +72,9 @@ void mlx4_cq_completion(struct mlx4_dev *dev, u32 cqn)
  	++cq->arm_sn;
 
  	cq->comp(cq);
+
+	if (atomic_dec_and_test(&cq->refcount))
+		complete(&cq->free);
  }
 
  void mlx4_cq_event(struct mlx4_dev *dev, u32 cqn, int event_type)

@@ -74,13 +82,13 @@ void mlx4_cq_event(struct mlx4_dev *dev, u32 cqn, int event_type)
  	struct mlx4_cq_table *cq_table = &mlx4_priv(dev)->cq_table;
  	struct mlx4_cq *cq;
 
-	spin_lock(&cq_table->lock);
+	rcu_read_lock();
 
  	cq = radix_tree_lookup(&cq_table->tree, cqn & (dev->caps.num_cqs - 1));
  	if (cq)
  		atomic_inc(&cq->refcount);
 
-	spin_unlock(&cq_table->lock);
+	rcu_read_unlock();
 
  	if (!cq) {
  		mlx4_warn(dev, "Async event for bogus CQ %08x\n", cqn);

@@ -328,6 +336,7 @@ err_radix:
  	spin_lock_irq(&cq_table->lock);
  	radix_tree_delete(&cq_table->tree, cq->cqn);
  	spin_unlock_irq(&cq_table->lock);
+	synchronize_rcu();
 
  err_icm:
  	mlx4_cq_free_icm(dev, cq->cqn);

My reading of the code (i.e. hw/mlx4/cq.c) is that the hardware cqe owner_sr_opcode field contains MLX4_CQE_OPCODE_ERROR when there is an error and therefore, the only way to recover what the opcode was is through the wr_id you used when submitting the WR.

Is my reading of the code correct?

Quoted 6 lines

My reading of the code (i.e. hw/mlx4/cq.c) is that the hardware cqe owner_sr_opcode field contains MLX4_CQE_OPCODE_ERROR when there is an error and therefore, the only way to recover what the opcode was is through the wr_id you used when submitting the WR.

Is my reading of the code correct?

I believe this is true wrt the IB spec.

I believe this is true wrt the IB spec.

===================================================
net/mlx4_core: Fix racy flow in the driver CQ completion handler

The mlx4 CQ completion handler, mlx4_cq_completion, doesn't bother to lock the radix tree which is used to manage the table of CQs, nor does it increase the reference count of the CQ before invoking the user provided callback (and decrease it afterwards).

This is racy and can cause use-after-free, null pointer dereference, etc, which result in kernel crashes.

To fix this, we must do the following in mlx4_cq_completion:
- increase the ref count on the cq before invoking the user callback, and decrement it after the callback.
- Place a lock around the radix tree lookup/ref-count-increase

Using an irq spinlock will not fix this issue. The problem is that under VPI, the ETH interface uses multiple msix irq's, which can result in one cq completion event interrupting another in-progress cq completion event. A deadlock results when the handler for the first cq completion grabs the spinlock, and is interrupted by the second completion before it has a chance to release the spinlock. The handler for the second completion will deadlock waiting for the spinlock to be released.

2. is possible in the Linux kernel for one hard irq callback to flash on CPU X while another hard irq callback is running on the same CPU?

Quoted 2 lines · Or Gerlitz

2. is possible in the Linux kernel for one hard irq callback to flash on CPU X while another hard irq callback is running on the same CPU?

I think that from kernel 2.6.35 on MSI IRQs are no longer nested. See also http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=753649dbc49345a73a2454c770a3f2d54d11aec6 or http://lwn.net/Articles/380931/

Quoted 8 lines · Bart Van Assche

Quoted 2 lines · Or Gerlitz

2. is possible in the Linux kernel for one hard irq callback to flash on CPU X while another hard irq callback is running on the same CPU?

I think that from kernel 2.6.35 on MSI IRQs are no longer nested. See also
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=753649dbc49345a73a2454c770a3f2d54d11aec6
or http://lwn.net/Articles/380931/

thanks, so suppose we agree on that, still the patch makes sense as the race is there, but does the patch has to change?

Or.

I just need to verify that the patch can be applied correctly on the upstream kernel. The use of RCU (and not spinlock) makes sense from a performance standpoint in any case. We do NOT want to force mlx4_cq_completion to have a spinlock which is device-global, resulting in having completion event processing be single-threaded in effect).