Re: Dropped packets mapping IRQs for adjusted queue counts on i40e

Re: Dropped packets mapping IRQs for adjusted queue counts on i40e

[Maciej Fijalkowski]

On Wed, May 05, 2021 at 01:01:28PM -0700, Jesse Brandeburg wrote:
> Zvi Effron wrote:
> 
> > On Tue, May 4, 2021 at 4:07 PM Zvi Effron <zeffron@riotgames.com> wrote:
> > > I'm suspecting it's something with how XDP_REDIRECT is implemented in
> > > the i40e driver, but I don't know if this is a) cross driver behavior,
> > > b) expected behavior, or c) a bug.
> > I think I've found the issue, and it appears to be specific to i40e
> > (and maybe other drivers, too, but not XDP itself).
> > 
> > When performing the XDP xmit, i40e uses the smp_processor_id() to
> > select the tx queue (see
> > https://elixir.bootlin.com/linux/v5.12.1/source/drivers/net/ethernet/intel/i40e/i40e_txrx.c#L3846).
> > I'm not 100% clear on how the CPU is selected (since we don't use
> > cores 0 and 1), we end up on a core whose id is higher than any
> > available queue.
> > 
> > I'm going to try to modify our IRQ mappings to test this.
> > 
> > If I'm correct, this feels like a bug to me, since it requires a user
> > to understand low level driver details to do IRQ remapping, which is a
> > bit higher level. But if it's intended, we'll just have to figure out
> > how to work around this. (Unfortunately, using split tx and rx queues
> > is not possible with i40e, so that easy solution is unavailable.)
> > 
> > --Zvi

Hey Zvi, sorry for the lack of assistance, there has been statutory free
time in Poland and today i'm in the birthday mode, but we managed to
discuss the issue with Magnus and we feel like we could have a solution
for that, more below.

> 
> 
> It seems like for Intel drivers, igc, ixgbe, i40e, ice all have
> this problem.
> 
> Notably, igb, fixes it like I would expect.

igb is correct but I think that we would like to avoid the introduction of
locking for higher speed NICs in XDP data path.

We talked with Magnus that for i40e and ice that have lots of HW
resources, we could always create the xdp_rings array of num_online_cpus()
size and use smp_processor_id() for accesses, regardless of the user's
changes to queue count.

This way the smp_processor_id() provides the serialization by itself as
we're under napi on a given cpu, so there's no need for locking
introduction - there is a per-cpu XDP ring provided. If we would stick to
the approach where you adjust the size of xdp_rings down to the shrinked
Rx queue count and use a smp_processor_id() % vsi->num_queue_pairs formula
then we could have a resource contention. Say that you did on a 16 core
system:
$ ethtool -L eth0 combined 2

and then mapped the q0 to cpu1 and q1 to cpu 11. Both queues will grab the
xdp_rings[1], so we would have to introduce the locking.

Proposed approach would just result with more Tx queues packed onto Tx
ring container of queue vector.

Thoughts? Any concerns? Should we have a 'fallback' mode if we would be
out of queues?

Currently I have a patch for ice (which is trivial) that does the thing
described above and I'm in the middle of testing it. I feel like this is
also addressing Jakub's input on the old patch that tried to resolve the
issue around that manner:

https://lore.kernel.org/netdev/20210123195219.55f6d4e1@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com/

> 
> Let's talk about it over on intel-wired-lan and cc netdev.

Let's do this now.

> 
> Jesse

Re: Dropped packets mapping IRQs for adjusted queue counts on i40e

[Toke Høiland-Jørgensen]

Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:

> On Wed, May 05, 2021 at 01:01:28PM -0700, Jesse Brandeburg wrote:
>> Zvi Effron wrote:
>> 
>> > On Tue, May 4, 2021 at 4:07 PM Zvi Effron <zeffron@riotgames.com> wrote:
>> > > I'm suspecting it's something with how XDP_REDIRECT is implemented in
>> > > the i40e driver, but I don't know if this is a) cross driver behavior,
>> > > b) expected behavior, or c) a bug.
>> > I think I've found the issue, and it appears to be specific to i40e
>> > (and maybe other drivers, too, but not XDP itself).
>> > 
>> > When performing the XDP xmit, i40e uses the smp_processor_id() to
>> > select the tx queue (see
>> > https://elixir.bootlin.com/linux/v5.12.1/source/drivers/net/ethernet/intel/i40e/i40e_txrx.c#L3846).
>> > I'm not 100% clear on how the CPU is selected (since we don't use
>> > cores 0 and 1), we end up on a core whose id is higher than any
>> > available queue.
>> > 
>> > I'm going to try to modify our IRQ mappings to test this.
>> > 
>> > If I'm correct, this feels like a bug to me, since it requires a user
>> > to understand low level driver details to do IRQ remapping, which is a
>> > bit higher level. But if it's intended, we'll just have to figure out
>> > how to work around this. (Unfortunately, using split tx and rx queues
>> > is not possible with i40e, so that easy solution is unavailable.)
>> > 
>> > --Zvi
>
> Hey Zvi, sorry for the lack of assistance, there has been statutory free
> time in Poland and today i'm in the birthday mode, but we managed to
> discuss the issue with Magnus and we feel like we could have a solution
> for that, more below.
>
>> 
>> 
>> It seems like for Intel drivers, igc, ixgbe, i40e, ice all have
>> this problem.
>> 
>> Notably, igb, fixes it like I would expect.
>
> igb is correct but I think that we would like to avoid the introduction of
> locking for higher speed NICs in XDP data path.
>
> We talked with Magnus that for i40e and ice that have lots of HW
> resources, we could always create the xdp_rings array of num_online_cpus()
> size and use smp_processor_id() for accesses, regardless of the user's
> changes to queue count.

What is "lots"? Systems with hundreds of CPUs exist (and I seem to
recall an issue with just such a system on Intel hardware(?)). Also,
what if num_online_cpus() changes?

> This way the smp_processor_id() provides the serialization by itself as
> we're under napi on a given cpu, so there's no need for locking
> introduction - there is a per-cpu XDP ring provided. If we would stick to
> the approach where you adjust the size of xdp_rings down to the shrinked
> Rx queue count and use a smp_processor_id() % vsi->num_queue_pairs formula
> then we could have a resource contention. Say that you did on a 16 core
> system:
> $ ethtool -L eth0 combined 2
>
> and then mapped the q0 to cpu1 and q1 to cpu 11. Both queues will grab the
> xdp_rings[1], so we would have to introduce the locking.
>
> Proposed approach would just result with more Tx queues packed onto Tx
> ring container of queue vector.
>
> Thoughts? Any concerns? Should we have a 'fallback' mode if we would be
> out of queues?

Yes, please :)

-Toke

Re: Dropped packets mapping IRQs for adjusted queue counts on i40e

[Maciej Fijalkowski]

On Thu, May 06, 2021 at 12:29:40PM +0200, Toke Høiland-Jørgensen wrote:
> Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:
> 
> > On Wed, May 05, 2021 at 01:01:28PM -0700, Jesse Brandeburg wrote:
> >> Zvi Effron wrote:
> >> 
> >> > On Tue, May 4, 2021 at 4:07 PM Zvi Effron <zeffron@riotgames.com> wrote:
> >> > > I'm suspecting it's something with how XDP_REDIRECT is implemented in
> >> > > the i40e driver, but I don't know if this is a) cross driver behavior,
> >> > > b) expected behavior, or c) a bug.
> >> > I think I've found the issue, and it appears to be specific to i40e
> >> > (and maybe other drivers, too, but not XDP itself).
> >> > 
> >> > When performing the XDP xmit, i40e uses the smp_processor_id() to
> >> > select the tx queue (see
> >> > https://elixir.bootlin.com/linux/v5.12.1/source/drivers/net/ethernet/intel/i40e/i40e_txrx.c#L3846).
> >> > I'm not 100% clear on how the CPU is selected (since we don't use
> >> > cores 0 and 1), we end up on a core whose id is higher than any
> >> > available queue.
> >> > 
> >> > I'm going to try to modify our IRQ mappings to test this.
> >> > 
> >> > If I'm correct, this feels like a bug to me, since it requires a user
> >> > to understand low level driver details to do IRQ remapping, which is a
> >> > bit higher level. But if it's intended, we'll just have to figure out
> >> > how to work around this. (Unfortunately, using split tx and rx queues
> >> > is not possible with i40e, so that easy solution is unavailable.)
> >> > 
> >> > --Zvi
> >
> > Hey Zvi, sorry for the lack of assistance, there has been statutory free
> > time in Poland and today i'm in the birthday mode, but we managed to
> > discuss the issue with Magnus and we feel like we could have a solution
> > for that, more below.
> >
> >> 
> >> 
> >> It seems like for Intel drivers, igc, ixgbe, i40e, ice all have
> >> this problem.
> >> 
> >> Notably, igb, fixes it like I would expect.
> >
> > igb is correct but I think that we would like to avoid the introduction of
> > locking for higher speed NICs in XDP data path.
> >
> > We talked with Magnus that for i40e and ice that have lots of HW
> > resources, we could always create the xdp_rings array of num_online_cpus()
> > size and use smp_processor_id() for accesses, regardless of the user's
> > changes to queue count.
> 
> What is "lots"? Systems with hundreds of CPUs exist (and I seem to
> recall an issue with just such a system on Intel hardware(?)). Also,
> what if num_online_cpus() changes?

"Lots" is 16k for ice. For i40e datasheet tells that it's only 1536 for
whole device, so I back off from the statement that i40e has a lot of
resources :)

Also, s/num_online_cpus()/num_possible_cpus().

> 
> > This way the smp_processor_id() provides the serialization by itself as
> > we're under napi on a given cpu, so there's no need for locking
> > introduction - there is a per-cpu XDP ring provided. If we would stick to
> > the approach where you adjust the size of xdp_rings down to the shrinked
> > Rx queue count and use a smp_processor_id() % vsi->num_queue_pairs formula
> > then we could have a resource contention. Say that you did on a 16 core
> > system:
> > $ ethtool -L eth0 combined 2
> >
> > and then mapped the q0 to cpu1 and q1 to cpu 11. Both queues will grab the
> > xdp_rings[1], so we would have to introduce the locking.
> >
> > Proposed approach would just result with more Tx queues packed onto Tx
> > ring container of queue vector.
> >
> > Thoughts? Any concerns? Should we have a 'fallback' mode if we would be
> > out of queues?
> 
> Yes, please :)

How to have a fallback (in drivers that need it) in a way that wouldn't
hurt the scenario where queue per cpu requirement is satisfied?

> 
> -Toke
> 

Re: Dropped packets mapping IRQs for adjusted queue counts on i40e

[Toke Høiland-Jørgensen]

Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:

> On Thu, May 06, 2021 at 12:29:40PM +0200, Toke Høiland-Jørgensen wrote:
>> Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:
>> 
>> > On Wed, May 05, 2021 at 01:01:28PM -0700, Jesse Brandeburg wrote:
>> >> Zvi Effron wrote:
>> >> 
>> >> > On Tue, May 4, 2021 at 4:07 PM Zvi Effron <zeffron@riotgames.com> wrote:
>> >> > > I'm suspecting it's something with how XDP_REDIRECT is implemented in
>> >> > > the i40e driver, but I don't know if this is a) cross driver behavior,
>> >> > > b) expected behavior, or c) a bug.
>> >> > I think I've found the issue, and it appears to be specific to i40e
>> >> > (and maybe other drivers, too, but not XDP itself).
>> >> > 
>> >> > When performing the XDP xmit, i40e uses the smp_processor_id() to
>> >> > select the tx queue (see
>> >> > https://elixir.bootlin.com/linux/v5.12.1/source/drivers/net/ethernet/intel/i40e/i40e_txrx.c#L3846).
>> >> > I'm not 100% clear on how the CPU is selected (since we don't use
>> >> > cores 0 and 1), we end up on a core whose id is higher than any
>> >> > available queue.
>> >> > 
>> >> > I'm going to try to modify our IRQ mappings to test this.
>> >> > 
>> >> > If I'm correct, this feels like a bug to me, since it requires a user
>> >> > to understand low level driver details to do IRQ remapping, which is a
>> >> > bit higher level. But if it's intended, we'll just have to figure out
>> >> > how to work around this. (Unfortunately, using split tx and rx queues
>> >> > is not possible with i40e, so that easy solution is unavailable.)
>> >> > 
>> >> > --Zvi
>> >
>> > Hey Zvi, sorry for the lack of assistance, there has been statutory free
>> > time in Poland and today i'm in the birthday mode, but we managed to
>> > discuss the issue with Magnus and we feel like we could have a solution
>> > for that, more below.
>> >
>> >> 
>> >> 
>> >> It seems like for Intel drivers, igc, ixgbe, i40e, ice all have
>> >> this problem.
>> >> 
>> >> Notably, igb, fixes it like I would expect.
>> >
>> > igb is correct but I think that we would like to avoid the introduction of
>> > locking for higher speed NICs in XDP data path.
>> >
>> > We talked with Magnus that for i40e and ice that have lots of HW
>> > resources, we could always create the xdp_rings array of num_online_cpus()
>> > size and use smp_processor_id() for accesses, regardless of the user's
>> > changes to queue count.
>> 
>> What is "lots"? Systems with hundreds of CPUs exist (and I seem to
>> recall an issue with just such a system on Intel hardware(?)). Also,
>> what if num_online_cpus() changes?
>
> "Lots" is 16k for ice. For i40e datasheet tells that it's only 1536 for
> whole device, so I back off from the statement that i40e has a lot of
> resources :)
>
> Also, s/num_online_cpus()/num_possible_cpus().

OK, even 1536 is more than I expected; I figured it would be way lower,
which is why you were suggesting to use num_online_cpus() instead; but
yeah, num_possible_cpus() is obviously better, then :)

>> > This way the smp_processor_id() provides the serialization by itself as
>> > we're under napi on a given cpu, so there's no need for locking
>> > introduction - there is a per-cpu XDP ring provided. If we would stick to
>> > the approach where you adjust the size of xdp_rings down to the shrinked
>> > Rx queue count and use a smp_processor_id() % vsi->num_queue_pairs formula
>> > then we could have a resource contention. Say that you did on a 16 core
>> > system:
>> > $ ethtool -L eth0 combined 2
>> >
>> > and then mapped the q0 to cpu1 and q1 to cpu 11. Both queues will grab the
>> > xdp_rings[1], so we would have to introduce the locking.
>> >
>> > Proposed approach would just result with more Tx queues packed onto Tx
>> > ring container of queue vector.
>> >
>> > Thoughts? Any concerns? Should we have a 'fallback' mode if we would be
>> > out of queues?
>> 
>> Yes, please :)
>
> How to have a fallback (in drivers that need it) in a way that wouldn't
> hurt the scenario where queue per cpu requirement is satisfied?

Well, it should be possible to detect this at setup time, right? Not too
familiar with the driver code, but would it be possible to statically
dispatch to an entirely different code path if this happens?

-Toke

Re: Dropped packets mapping IRQs for adjusted queue counts on i40e

[Jesper Dangaard Brouer]

On Mon, 10 May 2021 13:22:54 +0200
Toke Høiland-Jørgensen <toke@redhat.com> wrote:

> Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:
> 
> > On Thu, May 06, 2021 at 12:29:40PM +0200, Toke Høiland-Jørgensen wrote:  
> >> Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:
> >>   
> >> > On Wed, May 05, 2021 at 01:01:28PM -0700, Jesse Brandeburg wrote:  
> >> >> Zvi Effron wrote:
> >> >>   
> >> >> > On Tue, May 4, 2021 at 4:07 PM Zvi Effron <zeffron@riotgames.com> wrote:  
> >> >> > > I'm suspecting it's something with how XDP_REDIRECT is implemented in
> >> >> > > the i40e driver, but I don't know if this is a) cross driver behavior,
> >> >> > > b) expected behavior, or c) a bug.  
> >> >> > I think I've found the issue, and it appears to be specific to i40e
> >> >> > (and maybe other drivers, too, but not XDP itself).
> >> >> > 
> >> >> > When performing the XDP xmit, i40e uses the smp_processor_id() to
> >> >> > select the tx queue (see
> >> >> > https://elixir.bootlin.com/linux/v5.12.1/source/drivers/net/ethernet/intel/i40e/i40e_txrx.c#L3846).
> >> >> > I'm not 100% clear on how the CPU is selected (since we don't use
> >> >> > cores 0 and 1), we end up on a core whose id is higher than any
> >> >> > available queue.
> >> >> > 
> >> >> > I'm going to try to modify our IRQ mappings to test this.
> >> >> > 
> >> >> > If I'm correct, this feels like a bug to me, since it requires a user
> >> >> > to understand low level driver details to do IRQ remapping, which is a
> >> >> > bit higher level. But if it's intended, we'll just have to figure out
> >> >> > how to work around this. (Unfortunately, using split tx and rx queues
> >> >> > is not possible with i40e, so that easy solution is unavailable.)
> >> >> > 
> >> >> > --Zvi  
> >> >
> >> > Hey Zvi, sorry for the lack of assistance, there has been statutory free
> >> > time in Poland and today i'm in the birthday mode, but we managed to
> >> > discuss the issue with Magnus and we feel like we could have a solution
> >> > for that, more below.
> >> >  
> >> >> 
> >> >> 
> >> >> It seems like for Intel drivers, igc, ixgbe, i40e, ice all have
> >> >> this problem.
> >> >> 
> >> >> Notably, igb, fixes it like I would expect.  
> >> >
> >> > igb is correct but I think that we would like to avoid the introduction of
> >> > locking for higher speed NICs in XDP data path.
> >> >
> >> > We talked with Magnus that for i40e and ice that have lots of HW
> >> > resources, we could always create the xdp_rings array of num_online_cpus()
> >> > size and use smp_processor_id() for accesses, regardless of the user's
> >> > changes to queue count.  
> >> 
> >> What is "lots"? Systems with hundreds of CPUs exist (and I seem to
> >> recall an issue with just such a system on Intel hardware(?)). Also,
> >> what if num_online_cpus() changes?  
> >
> > "Lots" is 16k for ice. For i40e datasheet tells that it's only 1536 for
> > whole device, so I back off from the statement that i40e has a lot of
> > resources :)
> >
> > Also, s/num_online_cpus()/num_possible_cpus().  
> 
> OK, even 1536 is more than I expected; I figured it would be way lower,
> which is why you were suggesting to use num_online_cpus() instead; but
> yeah, num_possible_cpus() is obviously better, then :)
> 
> >> > This way the smp_processor_id() provides the serialization by itself as
> >> > we're under napi on a given cpu, so there's no need for locking
> >> > introduction - there is a per-cpu XDP ring provided. If we would stick to
> >> > the approach where you adjust the size of xdp_rings down to the shrinked
> >> > Rx queue count and use a smp_processor_id() % vsi->num_queue_pairs formula
> >> > then we could have a resource contention. Say that you did on a 16 core
> >> > system:
> >> > $ ethtool -L eth0 combined 2
> >> >
> >> > and then mapped the q0 to cpu1 and q1 to cpu 11. Both queues will grab the
> >> > xdp_rings[1], so we would have to introduce the locking.
> >> >
> >> > Proposed approach would just result with more Tx queues packed onto Tx
> >> > ring container of queue vector.
> >> >
> >> > Thoughts? Any concerns? Should we have a 'fallback' mode if we would be
> >> > out of queues?  
> >> 
> >> Yes, please :)  
> >
> > How to have a fallback (in drivers that need it) in a way that wouldn't
> > hurt the scenario where queue per cpu requirement is satisfied?  
> 
> Well, it should be possible to detect this at setup time, right? Not too
> familiar with the driver code, but would it be possible to statically
> dispatch to an entirely different code path if this happens?

The ndo_xdp_xmit call is a function pointer.  Thus, if it happens at
this level, then at setup time the driver can simply change the NDO to
use a TX-queue-locked variant.

I actually consider it a bug that i40e allow this misconfig to happen.
The ixgbe driver solves the problem by rejecting XDP attach if the
system have more CPUs than TXQs available.

IMHO it is a better solution to add shard'ed/partitioned TXQ-locking
when this situation happens, instead of denying XDP attach.  Since the
original XDP-redirect the ndo_xdp_xmit call have gotten bulking added,
thus the locking will be amortized over the bulk.

One question is how do we inform the end-user that XDP will be using
a slightly slower TXQ-locking scheme?  Given we have no XDP-features
exposed, I suggest a simple kernel log message, which we already have
for other XDP situations when the MTU is too large, or TSO is enabled.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer