[dpdk-dev,2/7] net/mlx4: fix Rx interrupts with multiple ports

Message ID 2407596e1e05260170f8b0b80806b28b63e9255d.1497439616.git.adrien.mazarguil@6wind.com (mailing list archive)
State Accepted, archived
Delegated to: Ferruh Yigit
Headers

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/Intel-compilation success Compilation OK

Commit Message

Adrien Mazarguil June 14, 2017, 11:49 a.m. UTC
  Several Ethernet device structures are allocated on top of a common PCI
device for mlx4 adapters with multiple ports. These inherit a common
interrupt handle from their parent PCI device, which prevents Rx interrupts
from working properly on all ports as their configuration is overwritten.

Use a local interrupt handle to address this issue.

Fixes: 9f05a4b81809 ("net/mlx4: support user space Rx interrupt event")

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Moti Haimovsky <motih@mellanox.com>
---
 drivers/net/mlx4/mlx4.c | 9 +++++++++
 drivers/net/mlx4/mlx4.h | 1 +
 2 files changed, 10 insertions(+)
  

Comments

Ferruh Yigit June 16, 2017, 1:07 p.m. UTC | #1
On 6/14/2017 12:49 PM, Adrien Mazarguil wrote:
> Several Ethernet device structures are allocated on top of a common PCI
> device for mlx4 adapters with multiple ports. These inherit a common
> interrupt handle from their parent PCI device, which prevents Rx interrupts
> from working properly on all ports as their configuration is overwritten.
> 
> Use a local interrupt handle to address this issue.

Hi Adrien,

I am not clear why local copy required, and main concern from my point
of view is if this is a common problem for all PMDs and should be
addressed in higher level?

The variable is already per eth_dev, but this patch moves it the private
data. What overwrites it within eth_dev?

Thanks,
ferruh

> 
> Fixes: 9f05a4b81809 ("net/mlx4: support user space Rx interrupt event")
> 
> Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
> Acked-by: Moti Haimovsky <motih@mellanox.com>
> ---
>  drivers/net/mlx4/mlx4.c | 9 +++++++++
>  drivers/net/mlx4/mlx4.h | 1 +
>  2 files changed, 10 insertions(+)
> 
> diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
> index 178562e..2b4722f 100644
> --- a/drivers/net/mlx4/mlx4.c
> +++ b/drivers/net/mlx4/mlx4.c
> @@ -6207,6 +6207,15 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
>  
>  		eth_dev->device->driver = &mlx4_driver.driver;
>  
> +		/*
> +		 * Copy and override interrupt handle to prevent it from
> +		 * being shared between all ethdev instances of a given PCI
> +		 * device. This is required to properly handle Rx interrupts
> +		 * on all ports.
> +		 */
> +		priv->intr_handle_dev = *eth_dev->intr_handle;
> +		eth_dev->intr_handle = &priv->intr_handle_dev;
> +
>  		priv->dev = eth_dev;
>  		eth_dev->dev_ops = &mlx4_dev_ops;
>  
> diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
> index c46fc23..b74fbf8 100644
> --- a/drivers/net/mlx4/mlx4.h
> +++ b/drivers/net/mlx4/mlx4.h
> @@ -345,6 +345,7 @@ struct priv {
>  	unsigned int txqs_n; /* TX queues array size. */
>  	struct rxq *(*rxqs)[]; /* RX queues. */
>  	struct txq *(*txqs)[]; /* TX queues. */
> +	struct rte_intr_handle intr_handle_dev; /* Device interrupt handler. */
>  	struct rte_intr_handle intr_handle; /* Interrupt handler. */
>  	struct rte_flow_drop *flow_drop_queue; /* Flow drop queue. */
>  	LIST_HEAD(mlx4_flows, rte_flow) flows;
>
  
Adrien Mazarguil June 16, 2017, 1:39 p.m. UTC | #2
On Fri, Jun 16, 2017 at 02:07:54PM +0100, Ferruh Yigit wrote:
> On 6/14/2017 12:49 PM, Adrien Mazarguil wrote:
> > Several Ethernet device structures are allocated on top of a common PCI
> > device for mlx4 adapters with multiple ports. These inherit a common
> > interrupt handle from their parent PCI device, which prevents Rx interrupts
> > from working properly on all ports as their configuration is overwritten.
> > 
> > Use a local interrupt handle to address this issue.
> 
> Hi Adrien,
> 
> I am not clear why local copy required, and main concern from my point
> of view is if this is a common problem for all PMDs and should be
> addressed in higher level?

This issue only affects PMDs that handle multiple Ethernet ports through a
single PCI device. Such PMDs (like mlx4) identify themselves as PCI drivers
that manually have to register multiple rte_eth_dev instances through
rte_eth_dev_allocate(), which they then have to initialize.

> The variable is already per eth_dev, but this patch moves it the private
> data. What overwrites it within eth_dev?

Calling rte_eth_copy_pci_info() makes the rte_eth_dev structure inherit the
default interrupt handle of the underlying PCI device. By "inherit", I mean
eth_dev->intr_handle points to it, in that sense it's not per eth_dev but
per PCI device.

mlx4 Rx interrupts are associated with a given Verbs context, and each port
has its own Verbs context, so they cannot be shared, while other PMDs using
other methods for catching interrupts may be perfectly fine with a single
vector associated with the PCI device. It depends on the PMD, for instance
there is no such problem with mlx5 as exactly one PCI device is associated
with a given port.

This patch merely allocates a specific interrupt handle associated with the
eth_dev itself and makes the eth_dev handle point to that instead of the
default PCI handle. This "local" handle is initialized using the PCI handle
as a template before modifying the pointer. It's completely safe.

> > Fixes: 9f05a4b81809 ("net/mlx4: support user space Rx interrupt event")
> > 
> > Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
> > Acked-by: Moti Haimovsky <motih@mellanox.com>
> > ---
> >  drivers/net/mlx4/mlx4.c | 9 +++++++++
> >  drivers/net/mlx4/mlx4.h | 1 +
> >  2 files changed, 10 insertions(+)
> > 
> > diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
> > index 178562e..2b4722f 100644
> > --- a/drivers/net/mlx4/mlx4.c
> > +++ b/drivers/net/mlx4/mlx4.c
> > @@ -6207,6 +6207,15 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
> >  
> >  		eth_dev->device->driver = &mlx4_driver.driver;
> >  
> > +		/*
> > +		 * Copy and override interrupt handle to prevent it from
> > +		 * being shared between all ethdev instances of a given PCI
> > +		 * device. This is required to properly handle Rx interrupts
> > +		 * on all ports.
> > +		 */
> > +		priv->intr_handle_dev = *eth_dev->intr_handle;
> > +		eth_dev->intr_handle = &priv->intr_handle_dev;
> > +
> >  		priv->dev = eth_dev;
> >  		eth_dev->dev_ops = &mlx4_dev_ops;
> >  
> > diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
> > index c46fc23..b74fbf8 100644
> > --- a/drivers/net/mlx4/mlx4.h
> > +++ b/drivers/net/mlx4/mlx4.h
> > @@ -345,6 +345,7 @@ struct priv {
> >  	unsigned int txqs_n; /* TX queues array size. */
> >  	struct rxq *(*rxqs)[]; /* RX queues. */
> >  	struct txq *(*txqs)[]; /* TX queues. */
> > +	struct rte_intr_handle intr_handle_dev; /* Device interrupt handler. */
> >  	struct rte_intr_handle intr_handle; /* Interrupt handler. */
> >  	struct rte_flow_drop *flow_drop_queue; /* Flow drop queue. */
> >  	LIST_HEAD(mlx4_flows, rte_flow) flows;
> > 
>
  
Ferruh Yigit June 16, 2017, 1:49 p.m. UTC | #3
On 6/16/2017 2:39 PM, Adrien Mazarguil wrote:
> On Fri, Jun 16, 2017 at 02:07:54PM +0100, Ferruh Yigit wrote:
>> On 6/14/2017 12:49 PM, Adrien Mazarguil wrote:
>>> Several Ethernet device structures are allocated on top of a common PCI
>>> device for mlx4 adapters with multiple ports. These inherit a common
>>> interrupt handle from their parent PCI device, which prevents Rx interrupts
>>> from working properly on all ports as their configuration is overwritten.
>>>
>>> Use a local interrupt handle to address this issue.
>>
>> Hi Adrien,
>>
>> I am not clear why local copy required, and main concern from my point
>> of view is if this is a common problem for all PMDs and should be
>> addressed in higher level?
> 
> This issue only affects PMDs that handle multiple Ethernet ports through a
> single PCI device. Such PMDs (like mlx4) identify themselves as PCI drivers
> that manually have to register multiple rte_eth_dev instances through
> rte_eth_dev_allocate(), which they then have to initialize.
> 
>> The variable is already per eth_dev, but this patch moves it the private
>> data. What overwrites it within eth_dev?
> 
> Calling rte_eth_copy_pci_info() makes the rte_eth_dev structure inherit the
> default interrupt handle of the underlying PCI device. By "inherit", I mean
> eth_dev->intr_handle points to it, in that sense it's not per eth_dev but
> per PCI device.
> 
> mlx4 Rx interrupts are associated with a given Verbs context, and each port
> has its own Verbs context, so they cannot be shared, while other PMDs using
> other methods for catching interrupts may be perfectly fine with a single
> vector associated with the PCI device. It depends on the PMD, for instance
> there is no such problem with mlx5 as exactly one PCI device is associated
> with a given port.
> 
> This patch merely allocates a specific interrupt handle associated with the
> eth_dev itself and makes the eth_dev handle point to that instead of the
> default PCI handle. This "local" handle is initialized using the PCI handle
> as a template before modifying the pointer. It's completely safe.

Thanks for clarification.

> 
>>> Fixes: 9f05a4b81809 ("net/mlx4: support user space Rx interrupt event")
>>>
>>> Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
>>> Acked-by: Moti Haimovsky <motih@mellanox.com>
>>> ---
>>>  drivers/net/mlx4/mlx4.c | 9 +++++++++
>>>  drivers/net/mlx4/mlx4.h | 1 +
>>>  2 files changed, 10 insertions(+)
>>>
>>> diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
>>> index 178562e..2b4722f 100644
>>> --- a/drivers/net/mlx4/mlx4.c
>>> +++ b/drivers/net/mlx4/mlx4.c
>>> @@ -6207,6 +6207,15 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
>>>  
>>>  		eth_dev->device->driver = &mlx4_driver.driver;
>>>  
>>> +		/*
>>> +		 * Copy and override interrupt handle to prevent it from
>>> +		 * being shared between all ethdev instances of a given PCI
>>> +		 * device. This is required to properly handle Rx interrupts
>>> +		 * on all ports.
>>> +		 */
>>> +		priv->intr_handle_dev = *eth_dev->intr_handle;
>>> +		eth_dev->intr_handle = &priv->intr_handle_dev;
>>> +
>>>  		priv->dev = eth_dev;
>>>  		eth_dev->dev_ops = &mlx4_dev_ops;
>>>  
>>> diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
>>> index c46fc23..b74fbf8 100644
>>> --- a/drivers/net/mlx4/mlx4.h
>>> +++ b/drivers/net/mlx4/mlx4.h
>>> @@ -345,6 +345,7 @@ struct priv {
>>>  	unsigned int txqs_n; /* TX queues array size. */
>>>  	struct rxq *(*rxqs)[]; /* RX queues. */
>>>  	struct txq *(*txqs)[]; /* TX queues. */
>>> +	struct rte_intr_handle intr_handle_dev; /* Device interrupt handler. */
>>>  	struct rte_intr_handle intr_handle; /* Interrupt handler. */
>>>  	struct rte_flow_drop *flow_drop_queue; /* Flow drop queue. */
>>>  	LIST_HEAD(mlx4_flows, rte_flow) flows;
>>>
>>
>
  

Patch

diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
index 178562e..2b4722f 100644
--- a/drivers/net/mlx4/mlx4.c
+++ b/drivers/net/mlx4/mlx4.c
@@ -6207,6 +6207,15 @@  mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 
 		eth_dev->device->driver = &mlx4_driver.driver;
 
+		/*
+		 * Copy and override interrupt handle to prevent it from
+		 * being shared between all ethdev instances of a given PCI
+		 * device. This is required to properly handle Rx interrupts
+		 * on all ports.
+		 */
+		priv->intr_handle_dev = *eth_dev->intr_handle;
+		eth_dev->intr_handle = &priv->intr_handle_dev;
+
 		priv->dev = eth_dev;
 		eth_dev->dev_ops = &mlx4_dev_ops;
 
diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
index c46fc23..b74fbf8 100644
--- a/drivers/net/mlx4/mlx4.h
+++ b/drivers/net/mlx4/mlx4.h
@@ -345,6 +345,7 @@  struct priv {
 	unsigned int txqs_n; /* TX queues array size. */
 	struct rxq *(*rxqs)[]; /* RX queues. */
 	struct txq *(*txqs)[]; /* TX queues. */
+	struct rte_intr_handle intr_handle_dev; /* Device interrupt handler. */
 	struct rte_intr_handle intr_handle; /* Interrupt handler. */
 	struct rte_flow_drop *flow_drop_queue; /* Flow drop queue. */
 	LIST_HEAD(mlx4_flows, rte_flow) flows;