Skip to contentSkip to conversation

mptcp: Optimize DSS fields in mptcp_options_received

PATCH In review Peter Krystad 4 participants 10 messages Jan 17, 2020 – Jan 22, 2020

Subsystems (2)
lore
Share & tools
lore

Peter Krystad

author
#0 lore raw

PATCH 1/1 mptcp: Optimize DSS fields in mptcp_options_received

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>
---
 include/linux/tcp.h | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +12 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};
 	u32	subflow_seq;
 	u16	data_len;
 	u8	mp_capable : 1,
Signature · 1 line
2.17.2
#1 lore raw
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

struct {
union {
struct {
u64 sndr_key;
u64 rcvr_key;
};
struct {
u64 data_ack;
u64 data_seq;
};
};
u32 data_ack32;
u32 data_seq32;
};

That does require some rearranging in later commits (reintroducing the top-level union to add the ipv6 ADD_ADDR data).

Quoted 5 lines
include/linux/tcp.h +0 -0
 	u32	subflow_seq;
 	u16	data_len;
 	u8	mp_capable : 1,
Signature · 1 line
2.17.2

--
Mat Martineau
Intel

#2 lore raw
Quoted 48 lines · Mat Martineau
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

(quoted diff) +0 -0
 	struct {
 		union {
 			struct {
 				u64	sndr_key;

Review

Quoted 10 lines
(quoted diff) +0 -0
 				u64	rcvr_key;
 			};
 			struct {
 				u64	data_ack;
 				u64	data_seq;
 			};
 		};
 		u32	data_ack32;
 		u32	data_seq32;
 	};

I'm really low on coffee here, so please perdom me for the likely dumb comment, but...

why do we need data_*32 and *_key simult? I think we don't allow MPC and DSS in the same packet ?!?

Thanks,

Paolo

author
#3 lore raw
Quoted 60 lines · Paolo Abeni
Quoted 48 lines · Mat Martineau
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

(quoted diff) +0 -0
 	struct {
 		union {
 			struct {
 				u64	sndr_key;

Review

Quoted 10 lines
(quoted diff) +0 -0
 				u64	rcvr_key;
 			};
 			struct {
 				u64	data_ack;
 				u64	data_seq;
 			};
 		};
 		u32	data_ack32;
 		u32	data_seq32;
 	};
Quoted 5 lines

I'm really low on coffee here, so please perdom me for the likely dumb comment, but...

why do we need data_*32 and *_key simult? I think we don't allow MPC and DSS in the same packet ?!?

The mutipath-tcp.org implementation would do this so (and also data_*32 + IPv4 add_addr). If we know we won't ever do these combinations, and we only want to work with ourselves, then it is not necessary.

Peter.

Quoted 4 lines

Thanks,

Paolo

#4 lore raw
Quoted 64 lines · Peter Krystad
Quoted 63 lines · Paolo Abeni
Quoted 48 lines · Mat Martineau
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

(quoted diff) +0 -0
 	struct {
 		union {
 			struct {
 				u64	sndr_key;

Review

Quoted 10 lines
(quoted diff) +0 -0
 				u64	rcvr_key;
 			};
 			struct {
 				u64	data_ack;
 				u64	data_seq;
 			};
 		};
 		u32	data_ack32;
 		u32	data_seq32;
 	};

I'm really low on coffee here, so please perdom me for the likely dumb comment, but...

Quoted 6 lines
Quoted 2 lines

why do we need data_*32 and *_key simult? I think we don't allow MPC and DSS in the same packet ?!?

The mutipath-tcp.org implementation would do this so (and also data_*32 + IPv4 add_addr). If we know we won't ever do these combinations, and we only want to work with ourselves, then it is not necessary.

Uhmm.. I *think* with MPTCP v1 MPC+DSS is no longer required (nor should be sent)? we have MPC+data instead...

(sorry I have the vague the feeling we already have discussed this in the past, but for sure I completely forgot the conclusion...)

Thanks,

Paolo

#5 lore raw

Hi Paolo, Peter, Mat,

Quoted 73 lines · Paolo Abeni
Quoted 64 lines · Peter Krystad
Quoted 63 lines · Paolo Abeni
Quoted 44 lines · Mat Martineau
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

Quoted 14 lines
  	struct {
  		union {
  			struct {
  				u64	sndr_key;
  				u64	rcvr_key;
  			};
  			struct {
  				u64	data_ack;
  				u64	data_seq;
  			};
  		};
  		u32	data_ack32;
  		u32	data_seq32;
  	};

I'm really low on coffee here, so please perdom me for the likely dumb comment, but...

Quoted 6 lines
Quoted 2 lines

why do we need data_*32 and *_key simult? I think we don't allow MPC and DSS in the same packet ?!?

The mutipath-tcp.org implementation would do this so (and also data_*32 + IPv4 add_addr). If we know we won't ever do these combinations, and we only want to work with ourselves, then it is not necessary.

Quoted 5 lines

Uhmm.. I *think* with MPTCP v1 MPC+DSS is no longer required (nor should be sent)? we have MPC+data instead...

(sorry I have the vague the feeling we already have discussed this in the past, but for sure I completely forgot the conclusion...)

Sorry to ask this question but do we want this change for part 2?

I understand the benefits and we should optimize mptcp_options_received structure at some points but from what I see, there are still some discussions about the content of the structure and having this union forces us to make sure we don't mix the usage of the members when parsing TCP options and when reading what has been parsed. Maybe it's a risk we don't want to take now for part 2.

We could postpone these optimizations for later before adding MP_JOIN and ADD_ADDR.

This patch and the "parent" one ("mptcp: Store 32 and 64-bit ACKs separately") are the two blocking ones to finalize part 2.

What do you think?

Cheers,
Matt

Signature · 5 lines
Matthieu Baerts | R&D Engineer
matthieu.baerts(a)tessares.net
Tessares SA | Hybrid Access Solutions
www.tessares.net
1 Avenue Jean Monnet, 1348 Louvain-la-Neuve, Belgium
author
#6 lore raw
Quoted 90 lines · Matthieu Baerts

Hi Paolo, Peter, Mat,

Quoted 72 lines · Paolo Abeni
Quoted 64 lines · Peter Krystad
Quoted 63 lines · Paolo Abeni
Quoted 44 lines · Mat Martineau
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

Quoted 14 lines
  	struct {
  		union {
  			struct {
  				u64	sndr_key;
  				u64	rcvr_key;
  			};
  			struct {
  				u64	data_ack;
  				u64	data_seq;
  			};
  		};
  		u32	data_ack32;
  		u32	data_seq32;
  	};

I'm really low on coffee here, so please perdom me for the likely dumb comment, but...

Quoted 6 lines
Quoted 2 lines

why do we need data_*32 and *_key simult? I think we don't allow MPC and DSS in the same packet ?!?

The mutipath-tcp.org implementation would do this so (and also data_*32 + IPv4 add_addr). If we know we won't ever do these combinations, and we only want to work with ourselves, then it is not necessary.

Quoted 5 lines

Uhmm.. I *think* with MPTCP v1 MPC+DSS is no longer required (nor should be sent)? we have MPC+data instead...

(sorry I have the vague the feeling we already have discussed this in the past, but for sure I completely forgot the conclusion...)

Sorry to ask this question but do we want this change for part 2?

I understand the benefits and we should optimize mptcp_options_received structure at some points but from what I see, there are still some discussions about the content of the structure and having this union forces us to make sure we don't mix the usage of the members when parsing TCP options and when reading what has been parsed. Maybe it's a risk we don't want to take now for part 2.

Quoted 7 lines

We could postpone these optimizations for later before adding MP_JOIN and ADD_ADDR.

This patch and the "parent" one ("mptcp: Store 32 and 64-bit ACKs separately") are the two blocking ones to finalize part 2.

What do you think?

I still think the optimization is worth addressing. Here is what I propose:

I will drop/withdraw the two commits regarding 32-bit ack/seq:
mptcp: Store 32 and 64-bit ACKs separately
mptcp: Re-factor expand_ack

Then I will re-submit the 'optimize' patches revised to reflect the dropping of the first two, adding checks to make sure use of the union is safe and to disallow simultaneous MPC+DSS. If we want to defer these until after v2 is submitted (to reduce risk and time to submission) that is also fine with me.

Peter.

Quoted 2 lines

Cheers,
Matt

#7 lore raw

Hi Peter,

Quoted 94 lines · Peter Krystad
Quoted 93 lines · Matthieu Baerts

Hi Paolo, Peter, Mat,

Quoted 72 lines · Paolo Abeni
Quoted 64 lines · Peter Krystad
Quoted 63 lines · Paolo Abeni
Quoted 44 lines · Mat Martineau
Quoted 26 lines · Peter Krystad

Use union to minimize space used in structure.
Allow 32-bit DSS to be received at the same time as MP_CAPABLE.

squashto: Implement MPTCP receive path

Signed-off-by Peter Krystad <peter.krystad(a)linux.intel.com>

---
include/linux/tcp.h | 18 ++++++++++++------

1 file changed, 12 insertions(+), 6 deletions(-)
include/linux/tcp.h +2 -6
index 6abf8d9f1a42..a9200a86a6d2 100644
@@ -80,12 +80,18 @@ struct tcp_sack_block {
 
 #if IS_ENABLED(CONFIG_MPTCP)
 struct mptcp_options_received {
-	u64	sndr_key;
-	u64	rcvr_key;
-	u32	data_ack32;
-	u32	data_seq32;
-	u64	data_ack;
-	u64	data_seq;
+	union {
+		struct {

include/linux/tcp.h struct tcp_sack_block { · line 94

Quoted 10 lines
include/linux/tcp.h +10 -0
+			u64	sndr_key;
+			u64	rcvr_key;
+			u32	data_ack32;
+			u32	data_seq32;
+		};
+		struct {
+			u64	data_ack;
+			u64	data_seq;
+		};
+	};

If one of the goals is to ensure that mixed 32/64 ack/seq is working (as mentioned in next commit message), I think it would be better to express that as a union with the data_ack/data_seq and sndr_key/rcvr_key values rather than depend on struct alignment:

Quoted 14 lines
   	struct {
   		union {
   			struct {
   				u64	sndr_key;
   				u64	rcvr_key;
   			};
   			struct {
   				u64	data_ack;
   				u64	data_seq;
   			};
   		};
   		u32	data_ack32;
   		u32	data_seq32;
   	};

I'm really low on coffee here, so please perdom me for the likely dumb comment, but...

Quoted 6 lines
Quoted 2 lines

why do we need data_*32 and *_key simult? I think we don't allow MPC and DSS in the same packet ?!?

The mutipath-tcp.org implementation would do this so (and also data_*32 + IPv4 add_addr). If we know we won't ever do these combinations, and we only want to work with ourselves, then it is not necessary.

Quoted 5 lines

Uhmm.. I *think* with MPTCP v1 MPC+DSS is no longer required (nor should be sent)? we have MPC+data instead...

(sorry I have the vague the feeling we already have discussed this in the past, but for sure I completely forgot the conclusion...)

Sorry to ask this question but do we want this change for part 2?

I understand the benefits and we should optimize mptcp_options_received structure at some points but from what I see, there are still some discussions about the content of the structure and having this union forces us to make sure we don't mix the usage of the members when parsing TCP options and when reading what has been parsed. Maybe it's a risk we don't want to take now for part 2.

We could postpone these optimizations for later before adding MP_JOIN and ADD_ADDR.

Quoted 6 lines
Quoted 4 lines

This patch and the "parent" one ("mptcp: Store 32 and 64-bit ACKs separately") are the two blocking ones to finalize part 2.

What do you think?

I still think the optimization is worth addressing.

Me too.

Quoted 11 lines

Here is what I propose:

I will drop/withdraw the two commits regarding 32-bit ack/seq:
mptcp: Store 32 and 64-bit ACKs separately
mptcp: Re-factor expand_ack

Then I will re-submit the 'optimize' patches revised to reflect the dropping of the first two, adding checks to make sure use of the union is safe and to disallow simultaneous MPC+DSS. If we want to defer these until after v2 is submitted (to reduce risk and time to submission) that is also fine with me.

I think it would be better to defer these optimizations.

For part 2 where we don't support MP_JOIN, ADD_ADDR, RM_ADDR, MP_FAIL, MP_PRIO and MP_FASTCLOSE, we can wait before adding complexity here. So in this patch-set with support of MP_CAPABLE (/!\ v0 and later v1) and DSS, I guess we can only save max 64bits per TCP socket which is nothing when you look at its current size.

If the maintainers wants us to already do this optimization job now, we can integrate what you are working on later. But at least we can have a review on the rest of the code as well :-)

Or do you think it will be mandatory to have this optimization now?

Cheers,
Matt

Signature · 5 lines
Matthieu Baerts | R&D Engineer
matthieu.baerts(a)tessares.net
Tessares SA | Hybrid Access Solutions
www.tessares.net
1 Avenue Jean Monnet, 1348 Louvain-la-Neuve, Belgium
author
#8 lore raw
Quoted 1 line · Matthieu Baerts

Hi Peter,

<snip>

Quoted 1 line

Or do you think it will be mandatory to have this optimization now?

Not mandatory, we can go as is. I realized (after some code review) just this morning that for v1 although we do not send/receive MPC+DSS we need to have space for both options at the same time. This fragment when processing MP_CAPABLE in mptcp_parse_option() shows the need:

	if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA) {
		/* Section 3.1.:
		 * "the data parameters in a MP_CAPABLE are semantically
		 * equivalent to those in a DSS option and can be used
		 * interchangeably."
		 */
		mp_opt->dss = 1;

The code for part 2 is fine as is. I'll send the optimizations for ADD_ADDR, etc. later.

Peter.

Quoted 2 lines

Cheers,
Matt

#9 lore raw

Hi Peter,

Quoted 12 lines · Peter Krystad
Quoted 1 line · Matthieu Baerts

Hi Peter,

<snip>

Quoted 1 line

Or do you think it will be mandatory to have this optimization now?

Not mandatory, we can go as is. I realized (after some code review) just this morning that for v1 although we do not send/receive MPC+DSS we need to have space for both options at the same time. This fragment when processing MP_CAPABLE in mptcp_parse_option() shows the need:

Quoted 7 lines
	if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA) {
		/* Section 3.1.:
		 * "the data parameters in a MP_CAPABLE are semantically
		 * equivalent to those in a DSS option and can be used
		 * interchangeably."
		 */
		mp_opt->dss = 1;

Good catch, we have to deal with v0 and v1. Then indeed better not to modify part 2 and do that in part 3.

Quoted 2 lines

The code for part 2 is fine as is. I'll send the optimizations for ADD_ADDR, etc. later.

Thank you!

Cheers,
Matt

Signature · 5 lines
Matthieu Baerts | R&D Engineer
matthieu.baerts(a)tessares.net
Tessares SA | Hybrid Access Solutions
www.tessares.net
1 Avenue Jean Monnet, 1348 Louvain-la-Neuve, Belgium