Re: MLX4 Cq Question
On 20/05/2013 17:53, Jack Morgenstein wrote: === 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. I am not sure to follow on two pieces here: 1. why we say that only mlx4_en uses multiple msix irq's? mlx4_ib also exposes multiple vectors (-- EQs -- MSI-X -- IRQ) and the iser driver use that, e.g creates multiple CQs each on different EQ 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? Or. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On 05/21/13 11:40, Or Gerlitz wrote: 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/. Bart. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On 21/05/2013 13:42, Bart Van Assche wrote: On 05/21/13 11:40, Or Gerlitz wrote: 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. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On Tuesday 21 May 2013 13:43, Or Gerlitz wrote: On 21/05/2013 13:42, Bart Van Assche wrote: On 05/21/13 11:40, Or Gerlitz wrote: 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). -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On 21/05/2013 17:13, Jack Morgenstein wrote: 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). cool, lets do that and re-submit -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On Saturday 18 May 2013 00:37, Roland Dreier wrote: On Fri, May 17, 2013 at 12:25 PM, Tom Tucker t...@opengridcomputing.com wrote: 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. === 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
Re: MLX4 Cq Question
On Mon, May 20, 2013 at 7:53 AM, Jack Morgenstein ja...@dev.mellanox.co.il wrote: This is racy and can cause use-after-free, null pointer dereference, etc, which result in kernel crashes. Sounds fine and I'd be happy to apply your final patch, but I'd be curious to know what the race is in more detail. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
Hi Guys, One other quick one. I've received conflicting claims on the validity of the wc.opcode when wc.status != 0 for mlx4 hardware. 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? Thanks, Tom On 5/20/13 9:53 AM, Jack Morgenstein wrote: On Saturday 18 May 2013 00:37, Roland Dreier wrote: On Fri, May 17, 2013 at 12:25 PM, Tom Tucker t...@opengridcomputing.com wrote: 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. === 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
RE: MLX4 Cq Question
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. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On 5/20/13 2:58 PM, Hefty, Sean wrote: 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. Thanks, this was my recollection as well. Tom -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On 18/05/2013 00:37, Roland Dreier wrote: 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). Jack, so do we finally agree to this analysis? last time when this was on the list, I was under the impression that there was no consensus and I also see that on the stack we provide to customers there's a patch of yours in that area, or it may fix another bug? Or. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
MLX4 Cq Question
Hi Roland, 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. Thanks, Tom -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: MLX4 Cq Question
On Fri, May 17, 2013 at 12:25 PM, Tom Tucker t...@opengridcomputing.com wrote: 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. -- To unsubscribe from this list: send the line unsubscribe linux-rdma in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html