Extend irq_set_affinity_notifier() to use a call chain

Hi,

I'm working for Mellanox on mlx4_en NIC driver.

We need to be able to be notified on irq affinity changes. This is because, during non-stop full bandwidth traffic, napi will poll constantly and no interrupt will be fired. Because of that, even if the user changes the irq affinity, polling will continue to be done on the original CPU that was chosen on the first packet.
We would like to be notified when the affinity is changed. When such an event happen, the driver will arm the interrupts and end the napi session. An interrupt will start a new napi session on the right CPU.

In order to do that, I need to add a new irq affinity notification callback (In addition to the existing cpu_rmap notification). For that I would like to extend irq_set_affinity_notifier() to have a notifier call-chain instead of a single notifier callback.

I wanted to hear your opinion on this, and unless there is a better solution, will send an RFC later on.

References:
- http://patchwork.ozlabs.org/patch/65244/ - Review done by Thomas Glexiber to Ben Hutchings first version of the irq affinity notifiers.
- http://patchwork.ozlabs.org/patch/79593/ - Final version of irq_set_affinity_notifier() that was applied

Thanks,
Amir

In order to do that, I need to add a new irq affinity notification callback (In addition to the existing cpu_rmap notification). For that I would like to extend irq_set_affinity_notifier() to have a notifier call-chain instead of a single notifier callback.

Quoted 4 lines · Amir Vadai

In order to do that, I need to add a new irq affinity notification callback (In addition to the existing cpu_rmap notification). For that I would like to extend irq_set_affinity_notifier() to have a notifier call-chain instead of a single notifier callback.

Why? "I would like" is a non argument.

Thanks,

tglx

Quoted 7 lines · Thomas Gleixner

Quoted 4 lines · Amir Vadai

In order to do that, I need to add a new irq affinity notification callback (In addition to the existing cpu_rmap notification). For that I would like to extend irq_set_affinity_notifier() to have a notifier call-chain instead of a single notifier callback.

Why? "I would like" is a non argument.

Current implementation enables only one callback to be registered for irq affinity change notifications.

cpu_rmap is registered be notified - for RFS purposes. mlx4_en (and probably other network drivers) needs to be notified too, in order to stop the napi polling on the old cpu and move to the new one. To enable more than 1 notification callbacks, I suggest to use a notifier call chain.

Quoted 11 lines · Amir Vadai

Quoted 7 lines · Thomas Gleixner

Quoted 4 lines · Amir Vadai

In order to do that, I need to add a new irq affinity notification callback (In addition to the existing cpu_rmap notification). For that I would like to extend irq_set_affinity_notifier() to have a notifier call-chain instead of a single notifier callback.

Why? "I would like" is a non argument.

Current implementation enables only one callback to be registered for irq affinity change notifications.

I'm well aware of that.

Quoted 5 lines

cpu_rmap is registered be notified - for RFS purposes. mlx4_en (and probably other network drivers) needs to be notified too, in order to stop the napi polling on the old cpu and move to the new one. To enable more than 1 notification callbacks, I suggest to use a notifier call chain.

You are not describing what needs to be notified and why. Please explain the details of that and how the RFS (whatever that is) and the network driver are connected

and why this notification cannot be
propagated inside the network stack itself.

notifier chains are almost always a clear sign for a design disaster and I'm not going to even think about it before I do not have a concice explanation of the problem at hand and why a notifier chain is a good solution.

Thanks,

tglx

Quoted 3 lines · Thomas Gleixner

You are not describing what needs to be notified and why. Please explain the details of that and how the RFS (whatever that is) and the network driver are connected

The goal of RFS is to increase datacache hitrate by steering kernel processing of packets in multi-queue devices to the CPU where the application thread consuming the packet is running.

In order to select the right queue, the networking stack needs to have a reverse map of IRQ affinty. This is the rmap that was added by Ben Hutchings [1]. To keep the rmap updated, cpu_rmap registers on the affinity notify.

This is the first affinity callback - it is located as a general library and not under net/...

The motivation to the second irq affinity callback is:
When traffic starts, first packet fires an interrupt which starts the napi polling on the cpu according the irq affinity.
If there is always packets to be consumed by the napi polling, no further interrupts will be fired, and napi will consume all the packets from the cpu it was started.
If the user changes the irq affinity, napi polling will continue to be done from the original cpu.
Only when the traffic will pause, napi session will be finished, and when traffic will resume, the new napi session will be done from the new cpu. This is a problematic behavior, because from the user point of view, cpu affinity can't be changed in a non-stop traffic scenario.

To solve this, the network driver should be notified on irq affinity change event, and restart the napi session. This could be done by closing the napi session and arming the interrupts. Next packet arrives will trigger an interrupt and napi will session will start, this time on the new CPU.

Quoted 2 lines

and why this notification cannot be
propagated inside the network stack itself.

To my understanding, those are two different consumers to the same event, one is a general library to maintain a reverse irq affinity map, and the other is networking specific, and maybe even a networking driver specific.

The rmap _IS_ instantiated by the driver, and both the driver and the networking core know about it.

So it's not completely different consumers. Just because it's a library does not mean it's disjunct from the code which uses it.

Aside of the fact, that maintaining a per irq notifier chain is going to be ugly as hell due to life time and locking issues, it's just opening a can of worms. How do you make sure that the invocation order is correct? What are the dependency rules of the driver restarting the napi session versus updating the rmap?

Even if you'd solve that and have a callback in the driver, then the callback never can restart the napi session directly. All it can do is set a flag which needs to be checked in the RX path, right?

So what's the point of adding notifier call chain complexity, ordering problems etc., if you can simply note the fact that the affinity changed in the rmap itself and check that in the RX path?

Thanks,

tglx

Quoted 4 lines · Thomas Gleixner

Even if you'd solve that and have a callback in the driver, then the callback never can restart the napi session directly. All it can do is set a flag which needs to be checked in the RX path, right?

Quoted 3 lines

So what's the point of adding notifier call chain complexity, ordering problems etc., if you can simply note the fact that the affinity changed in the rmap itself and check that in the RX path?

I will try to find a solution in the spirit of what you suggested - to let the rmap library notify napi about affinity changes - without adding this complexity to the code.