For skbedit action queue_mapping to select the transmit queue,
queue_mapping takes the value N for tx-N (where N is the actual
queue number). However, current behavior is the following:
1. Queue selection is off by 1, tx queue N-1 is selected for
action skbedit queue_mapping N. (If the general syntax for queue
index is 1 based, i.e., action skbedit queue_mapping N would
transmit to tx queue N-1, where N >=1, then the last queue cannot
be used for transmit as this fails the upper bound check.)
2. Transmit to first queue of TCs other than TC0 selects the
next queue.
3. It is not possible to transmit to the first queue (tx-0) as
this fails the bounds check, in this case the fallback
mechanism for hash calculation is used.
Fix the call to skb_set_queue_mapping(), the code retrieving the
transmit queue uses skb_get_rx_queue() which subtracts the queue
index by 1. This makes it so that "action skbedit queue_mapping N"
will transmit to tx-N (including the first and last queue).
Signed-off-by: Amritha Nambiar <redacted>
Reviewed-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
net/sched/act_skbedit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
From: Alexander Duyck <hidden> Date: 2021-08-10 00:51:25
On Mon, Aug 9, 2021 at 4:36 PM Amritha Nambiar
[off-list ref] wrote:
quoted hunk
For skbedit action queue_mapping to select the transmit queue,
queue_mapping takes the value N for tx-N (where N is the actual
queue number). However, current behavior is the following:
1. Queue selection is off by 1, tx queue N-1 is selected for
action skbedit queue_mapping N. (If the general syntax for queue
index is 1 based, i.e., action skbedit queue_mapping N would
transmit to tx queue N-1, where N >=1, then the last queue cannot
be used for transmit as this fails the upper bound check.)
2. Transmit to first queue of TCs other than TC0 selects the
next queue.
3. It is not possible to transmit to the first queue (tx-0) as
this fails the bounds check, in this case the fallback
mechanism for hash calculation is used.
Fix the call to skb_set_queue_mapping(), the code retrieving the
transmit queue uses skb_get_rx_queue() which subtracts the queue
index by 1. This makes it so that "action skbedit queue_mapping N"
will transmit to tx-N (including the first and last queue).
Signed-off-by: Amritha Nambiar <redacted>
Reviewed-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
net/sched/act_skbedit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
I don't think this is correct. It is conflating the rx_queue_mapping
versus the Tx queue mapping. This is supposed to be setting the Tx
queue mapping which applies after we have dropped the Rx queue
mapping, not before. Specifically this is run at the qdisc enqueue
stage with a single locked qdisc, after netdev_pick_tx and skb_tx_hash
have already run. It is something that existed before mq and is meant
to work with the mutliq qdisc.
If you are wanting to add a seperate override to add support for
programming the Rx queue mapping you may want to submit that as a
different patch rather than trying to change the existing Tx queue
mapping feature. Either that or you would need to change this so that
it has a different behavior depending on where the hook is added since
the behavior would be different if this is called before skb_tx_hash.
-----Original Message-----
From: Alexander Duyck <redacted>
Sent: Monday, August 9, 2021 5:51 PM
To: Nambiar, Amritha <redacted>
Cc: Netdev <redacted>; Jakub Kicinski <kuba@kernel.org>;
Jamal Hadi Salim [off-list ref]; Jiri Pirko [off-list ref]; Cong
Wang [off-list ref]; Samudrala, Sridhar
[off-list ref]
Subject: Re: [net-next PATCH] net: act_skbedit: Fix tc action skbedit
queue_mapping
On Mon, Aug 9, 2021 at 4:36 PM Amritha Nambiar
[off-list ref] wrote:
quoted
For skbedit action queue_mapping to select the transmit queue,
queue_mapping takes the value N for tx-N (where N is the actual
queue number). However, current behavior is the following:
1. Queue selection is off by 1, tx queue N-1 is selected for
action skbedit queue_mapping N. (If the general syntax for queue
index is 1 based, i.e., action skbedit queue_mapping N would
transmit to tx queue N-1, where N >=1, then the last queue cannot
be used for transmit as this fails the upper bound check.)
2. Transmit to first queue of TCs other than TC0 selects the
next queue.
3. It is not possible to transmit to the first queue (tx-0) as
this fails the bounds check, in this case the fallback
mechanism for hash calculation is used.
Fix the call to skb_set_queue_mapping(), the code retrieving the
transmit queue uses skb_get_rx_queue() which subtracts the queue
index by 1. This makes it so that "action skbedit queue_mapping N"
will transmit to tx-N (including the first and last queue).
Signed-off-by: Amritha Nambiar <redacted>
Reviewed-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
net/sched/act_skbedit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
I don't think this is correct. It is conflating the rx_queue_mapping
versus the Tx queue mapping. This is supposed to be setting the Tx
queue mapping which applies after we have dropped the Rx queue
mapping, not before. Specifically this is run at the qdisc enqueue
stage with a single locked qdisc, after netdev_pick_tx and skb_tx_hash
have already run. It is something that existed before mq and is meant
to work with the mutliq qdisc.
If you are wanting to add a seperate override to add support for
programming the Rx queue mapping you may want to submit that as a
different patch rather than trying to change the existing Tx queue
mapping feature. Either that or you would need to change this so that
it has a different behavior depending on where the hook is added since
the behavior would be different if this is called before skb_tx_hash.
Hi Alex,
Thanks for the review. The goal is to select a transmit queue using tc egress rule
and the action skbedit (that will go through netdev_pick_tx and skb_tx_hash).
I am not sure of the correct syntax for the queue-mapping value in the
action (tx-N or tx-N+1). As per the man page
(https://man7.org/linux/man-pages/man8/tc-skbedit.8.html), I interpreted
it as "action skbedit queue_mapping N" will transmit to tx-N. But, the
3 observations I listed don't quite seem to be following the tc rule.
Hence, tried to fix this in the action module.
-Amritha
From: Alexander Duyck <hidden> Date: 2021-08-10 14:04:01
On Mon, Aug 9, 2021 at 6:20 PM Nambiar, Amritha
[off-list ref] wrote:
quoted
-----Original Message-----
From: Alexander Duyck <redacted>
Sent: Monday, August 9, 2021 5:51 PM
To: Nambiar, Amritha <redacted>
Cc: Netdev <redacted>; Jakub Kicinski <kuba@kernel.org>;
Jamal Hadi Salim [off-list ref]; Jiri Pirko [off-list ref]; Cong
Wang [off-list ref]; Samudrala, Sridhar
[off-list ref]
Subject: Re: [net-next PATCH] net: act_skbedit: Fix tc action skbedit
queue_mapping
On Mon, Aug 9, 2021 at 4:36 PM Amritha Nambiar
[off-list ref] wrote:
quoted
For skbedit action queue_mapping to select the transmit queue,
queue_mapping takes the value N for tx-N (where N is the actual
queue number). However, current behavior is the following:
1. Queue selection is off by 1, tx queue N-1 is selected for
action skbedit queue_mapping N. (If the general syntax for queue
index is 1 based, i.e., action skbedit queue_mapping N would
transmit to tx queue N-1, where N >=1, then the last queue cannot
be used for transmit as this fails the upper bound check.)
2. Transmit to first queue of TCs other than TC0 selects the
next queue.
3. It is not possible to transmit to the first queue (tx-0) as
this fails the bounds check, in this case the fallback
mechanism for hash calculation is used.
Fix the call to skb_set_queue_mapping(), the code retrieving the
transmit queue uses skb_get_rx_queue() which subtracts the queue
index by 1. This makes it so that "action skbedit queue_mapping N"
will transmit to tx-N (including the first and last queue).
Signed-off-by: Amritha Nambiar <redacted>
Reviewed-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
net/sched/act_skbedit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
I don't think this is correct. It is conflating the rx_queue_mapping
versus the Tx queue mapping. This is supposed to be setting the Tx
queue mapping which applies after we have dropped the Rx queue
mapping, not before. Specifically this is run at the qdisc enqueue
stage with a single locked qdisc, after netdev_pick_tx and skb_tx_hash
have already run. It is something that existed before mq and is meant
to work with the mutliq qdisc.
If you are wanting to add a seperate override to add support for
programming the Rx queue mapping you may want to submit that as a
different patch rather than trying to change the existing Tx queue
mapping feature. Either that or you would need to change this so that
it has a different behavior depending on where the hook is added since
the behavior would be different if this is called before skb_tx_hash.
Hi Alex,
Thanks for the review. The goal is to select a transmit queue using tc egress rule
and the action skbedit (that will go through netdev_pick_tx and skb_tx_hash).
I am not sure of the correct syntax for the queue-mapping value in the
action (tx-N or tx-N+1). As per the man page
(https://man7.org/linux/man-pages/man8/tc-skbedit.8.html), I interpreted
it as "action skbedit queue_mapping N" will transmit to tx-N. But, the
3 observations I listed don't quite seem to be following the tc rule.
Hence, tried to fix this in the action module.
-Amritha
Hi Amritha,
As I mentioned before the problem is where the hook is being inserted.
If you follow the example in the documentation found at:
https://elixir.bootlin.com/linux/v5.14-rc5/source/Documentation/networking/multiqueue.rst
What you should find is that the Tx queue hook works as expected
because it will occur when the packet is enqueued to the qdisc, not
before. The problem is for the mq qdiscs this is occuring before the
Tx queue selection where the Rx queue is still active and the Tx queue
has not yet been selected.
You might take a look at the spots where tcf_classify is called. The
problem you are experiencing is because this action is expecting to
edit the packet inside the qdisc, whereas what you are doing is
calling this in sch_handle_egress which occurs before the Tx queue has
been selected.
What we likely need is some way of identifying when we are attaching
the classifier/action to either the ingress or egress classifier
rather than a qdisc and in that case then we use the
skb_record_rx_queue functionality instead of the skb_set_queue_mapping
functionality. One possible way to address this would be to expand
tc_at_ingress to actually be a 2 bit value and have 0 indicate Qdisc,
1 indicate Ingress, and 2 indicate Egress. Then you could filter out
the Ingress/Egress cases and have them use skb_record_rx_queue while
the other cases use skb_set_queue_mapping.
Hope that helps to clear it up a bit.
- Alex
-----Original Message-----
From: Alexander Duyck <redacted>
Sent: Tuesday, August 10, 2021 7:04 AM
To: Nambiar, Amritha <redacted>
Cc: Netdev <redacted>; Jakub Kicinski <kuba@kernel.org>;
Jamal Hadi Salim [off-list ref]; Jiri Pirko [off-list ref]; Cong
Wang [off-list ref]; Samudrala, Sridhar
[off-list ref]
Subject: Re: [net-next PATCH] net: act_skbedit: Fix tc action skbedit
queue_mapping
On Mon, Aug 9, 2021 at 6:20 PM Nambiar, Amritha
[off-list ref] wrote:
quoted
quoted
-----Original Message-----
From: Alexander Duyck <redacted>
Sent: Monday, August 9, 2021 5:51 PM
To: Nambiar, Amritha <redacted>
Cc: Netdev <redacted>; Jakub Kicinski <kuba@kernel.org>;
Jamal Hadi Salim [off-list ref]; Jiri Pirko [off-list ref]; Cong
Wang [off-list ref]; Samudrala, Sridhar
[off-list ref]
Subject: Re: [net-next PATCH] net: act_skbedit: Fix tc action skbedit
queue_mapping
On Mon, Aug 9, 2021 at 4:36 PM Amritha Nambiar
[off-list ref] wrote:
quoted
For skbedit action queue_mapping to select the transmit queue,
queue_mapping takes the value N for tx-N (where N is the actual
queue number). However, current behavior is the following:
1. Queue selection is off by 1, tx queue N-1 is selected for
action skbedit queue_mapping N. (If the general syntax for queue
index is 1 based, i.e., action skbedit queue_mapping N would
transmit to tx queue N-1, where N >=1, then the last queue cannot
be used for transmit as this fails the upper bound check.)
2. Transmit to first queue of TCs other than TC0 selects the
next queue.
3. It is not possible to transmit to the first queue (tx-0) as
this fails the bounds check, in this case the fallback
mechanism for hash calculation is used.
Fix the call to skb_set_queue_mapping(), the code retrieving the
transmit queue uses skb_get_rx_queue() which subtracts the queue
index by 1. This makes it so that "action skbedit queue_mapping N"
will transmit to tx-N (including the first and last queue).
Signed-off-by: Amritha Nambiar <redacted>
Reviewed-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
net/sched/act_skbedit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
I don't think this is correct. It is conflating the rx_queue_mapping
versus the Tx queue mapping. This is supposed to be setting the Tx
queue mapping which applies after we have dropped the Rx queue
mapping, not before. Specifically this is run at the qdisc enqueue
stage with a single locked qdisc, after netdev_pick_tx and skb_tx_hash
have already run. It is something that existed before mq and is meant
to work with the mutliq qdisc.
If you are wanting to add a seperate override to add support for
programming the Rx queue mapping you may want to submit that as a
different patch rather than trying to change the existing Tx queue
mapping feature. Either that or you would need to change this so that
it has a different behavior depending on where the hook is added since
the behavior would be different if this is called before skb_tx_hash.
Hi Alex,
Thanks for the review. The goal is to select a transmit queue using tc egress
rule
quoted
and the action skbedit (that will go through netdev_pick_tx and
skb_tx_hash).
quoted
I am not sure of the correct syntax for the queue-mapping value in the
action (tx-N or tx-N+1). As per the man page
(https://man7.org/linux/man-pages/man8/tc-skbedit.8.html), I interpreted
it as "action skbedit queue_mapping N" will transmit to tx-N. But, the
3 observations I listed don't quite seem to be following the tc rule.
Hence, tried to fix this in the action module.
-Amritha
Hi Amritha,
As I mentioned before the problem is where the hook is being inserted.
If you follow the example in the documentation found at:
https://elixir.bootlin.com/linux/v5.14-
rc5/source/Documentation/networking/multiqueue.rst
What you should find is that the Tx queue hook works as expected
because it will occur when the packet is enqueued to the qdisc, not
before. The problem is for the mq qdiscs this is occuring before the
Tx queue selection where the Rx queue is still active and the Tx queue
has not yet been selected.
You might take a look at the spots where tcf_classify is called. The
problem you are experiencing is because this action is expecting to
edit the packet inside the qdisc, whereas what you are doing is
calling this in sch_handle_egress which occurs before the Tx queue has
been selected.
What we likely need is some way of identifying when we are attaching
the classifier/action to either the ingress or egress classifier
rather than a qdisc and in that case then we use the
skb_record_rx_queue functionality instead of the skb_set_queue_mapping
functionality. One possible way to address this would be to expand
tc_at_ingress to actually be a 2 bit value and have 0 indicate Qdisc,
1 indicate Ingress, and 2 indicate Egress. Then you could filter out
the Ingress/Egress cases and have them use skb_record_rx_queue while
the other cases use skb_set_queue_mapping.
Hope that helps to clear it up a bit.
- Alex
Thanks for clarifying. Yes, it works as expected with the multiq qdisc. So, IIUC,
with multiq qdisc, this is hooked such that, netdev_pick_tx and skb_tx_hash
first picks a Tx queue. Then multiq_classify overwrites the queue_mapping
with the value from the skbedit action and calls qdisc_enqueue.
Yes, as you explained, I was hooking this up at the classifier level. I was adding an
egress filter to the clsact qdisc, and expecting the Tx queue to be picked from the
action when netdev_pick_tx executes from sch_handle_egress. This was not getting
into qdisc enqueue stage.
What I actually hoped to do was to add similar functionality (as with multiq) to
other qdiscs such as mqprio. So, to support this action with mqprio qdisc,
I would need to implement the enqueue callback in mqprio and set the
queue_mapping from the action, then run qdisc_enqueue. Is this correct ?
-Amritha
From: Alexander Duyck <hidden> Date: 2021-08-18 00:52:22
On Tue, Aug 17, 2021 at 5:16 PM Nambiar, Amritha
[off-list ref] wrote:
quoted
-----Original Message-----
From: Alexander Duyck <redacted>
Sent: Tuesday, August 10, 2021 7:04 AM
To: Nambiar, Amritha <redacted>
Cc: Netdev <redacted>; Jakub Kicinski <kuba@kernel.org>;
Jamal Hadi Salim [off-list ref]; Jiri Pirko [off-list ref]; Cong
Wang [off-list ref]; Samudrala, Sridhar
[off-list ref]
Subject: Re: [net-next PATCH] net: act_skbedit: Fix tc action skbedit
queue_mapping
On Mon, Aug 9, 2021 at 6:20 PM Nambiar, Amritha
[off-list ref] wrote:
quoted
quoted
-----Original Message-----
From: Alexander Duyck <redacted>
Sent: Monday, August 9, 2021 5:51 PM
To: Nambiar, Amritha <redacted>
Cc: Netdev <redacted>; Jakub Kicinski <kuba@kernel.org>;
Jamal Hadi Salim [off-list ref]; Jiri Pirko [off-list ref]; Cong
Wang [off-list ref]; Samudrala, Sridhar
[off-list ref]
Subject: Re: [net-next PATCH] net: act_skbedit: Fix tc action skbedit
queue_mapping
On Mon, Aug 9, 2021 at 4:36 PM Amritha Nambiar
[off-list ref] wrote:
quoted
For skbedit action queue_mapping to select the transmit queue,
queue_mapping takes the value N for tx-N (where N is the actual
queue number). However, current behavior is the following:
1. Queue selection is off by 1, tx queue N-1 is selected for
action skbedit queue_mapping N. (If the general syntax for queue
index is 1 based, i.e., action skbedit queue_mapping N would
transmit to tx queue N-1, where N >=1, then the last queue cannot
be used for transmit as this fails the upper bound check.)
2. Transmit to first queue of TCs other than TC0 selects the
next queue.
3. It is not possible to transmit to the first queue (tx-0) as
this fails the bounds check, in this case the fallback
mechanism for hash calculation is used.
Fix the call to skb_set_queue_mapping(), the code retrieving the
transmit queue uses skb_get_rx_queue() which subtracts the queue
index by 1. This makes it so that "action skbedit queue_mapping N"
will transmit to tx-N (including the first and last queue).
Signed-off-by: Amritha Nambiar <redacted>
Reviewed-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
net/sched/act_skbedit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
I don't think this is correct. It is conflating the rx_queue_mapping
versus the Tx queue mapping. This is supposed to be setting the Tx
queue mapping which applies after we have dropped the Rx queue
mapping, not before. Specifically this is run at the qdisc enqueue
stage with a single locked qdisc, after netdev_pick_tx and skb_tx_hash
have already run. It is something that existed before mq and is meant
to work with the mutliq qdisc.
If you are wanting to add a seperate override to add support for
programming the Rx queue mapping you may want to submit that as a
different patch rather than trying to change the existing Tx queue
mapping feature. Either that or you would need to change this so that
it has a different behavior depending on where the hook is added since
the behavior would be different if this is called before skb_tx_hash.
Hi Alex,
Thanks for the review. The goal is to select a transmit queue using tc egress
rule
quoted
and the action skbedit (that will go through netdev_pick_tx and
skb_tx_hash).
quoted
I am not sure of the correct syntax for the queue-mapping value in the
action (tx-N or tx-N+1). As per the man page
(https://man7.org/linux/man-pages/man8/tc-skbedit.8.html), I interpreted
it as "action skbedit queue_mapping N" will transmit to tx-N. But, the
3 observations I listed don't quite seem to be following the tc rule.
Hence, tried to fix this in the action module.
-Amritha
Hi Amritha,
As I mentioned before the problem is where the hook is being inserted.
If you follow the example in the documentation found at:
https://elixir.bootlin.com/linux/v5.14-
rc5/source/Documentation/networking/multiqueue.rst
What you should find is that the Tx queue hook works as expected
because it will occur when the packet is enqueued to the qdisc, not
before. The problem is for the mq qdiscs this is occuring before the
Tx queue selection where the Rx queue is still active and the Tx queue
has not yet been selected.
You might take a look at the spots where tcf_classify is called. The
problem you are experiencing is because this action is expecting to
edit the packet inside the qdisc, whereas what you are doing is
calling this in sch_handle_egress which occurs before the Tx queue has
been selected.
What we likely need is some way of identifying when we are attaching
the classifier/action to either the ingress or egress classifier
rather than a qdisc and in that case then we use the
skb_record_rx_queue functionality instead of the skb_set_queue_mapping
functionality. One possible way to address this would be to expand
tc_at_ingress to actually be a 2 bit value and have 0 indicate Qdisc,
1 indicate Ingress, and 2 indicate Egress. Then you could filter out
the Ingress/Egress cases and have them use skb_record_rx_queue while
the other cases use skb_set_queue_mapping.
Hope that helps to clear it up a bit.
- Alex
Thanks for clarifying. Yes, it works as expected with the multiq qdisc. So, IIUC,
with multiq qdisc, this is hooked such that, netdev_pick_tx and skb_tx_hash
first picks a Tx queue. Then multiq_classify overwrites the queue_mapping
with the value from the skbedit action and calls qdisc_enqueue.
Yes, as you explained, I was hooking this up at the classifier level. I was adding an
egress filter to the clsact qdisc, and expecting the Tx queue to be picked from the
action when netdev_pick_tx executes from sch_handle_egress. This was not getting
into qdisc enqueue stage.
What I actually hoped to do was to add similar functionality (as with multiq) to
other qdiscs such as mqprio. So, to support this action with mqprio qdisc,
I would need to implement the enqueue callback in mqprio and set the
queue_mapping from the action, then run qdisc_enqueue. Is this correct ?
So the problem isn't really with the qdiscs. The problem is the
ingress and egress hooks.
What you effectively have is one behavior you want for if this is
being handled on a qdisc, and another if you are adding it at the
ingress, sch_handle_ingress(), or egress, sch_handle_egress(), hook.
The qdisc behavior is as expected and really we wouldn't want to be
using mqprio because it forces us to serialize all the packets to use
it at that stage.
That is why I was thinking it might make sense to look at expanding
tc_at_ingress so it could become a 2 bit enum. Then you use 0 as the
qdisc value, 1 as the egress, and 2 as the ingress value. With that
you could then simply take the Rx queue recording approach for if the
value is set, and the Tx approach for if it is not. The overall change
wouldn't be too big as it would mostly be shifting a write or two
around, doing a rename, and maybe adding a compare in act_mirred.c
Thanks,
- Alex