Skip to contentSkip to conversation

gro: some minor optimizations

RFC PATCH In review Paolo Abeni 3 participants 10 messages Jan 18, 2022 – Feb 7, 2022

Subsystems (2)
lore
Share & tools
lore

Paolo Abeni

author maintainer
#0 lore raw

cover gro: some minor optimizations

This series collects a few small optimization for the GRO engine. I measure a 10% performance improvements in micro-benchmarks around dev_gro_receive(), but deltas are within noise range in tput tests - ence the RFC tag.

Any feedback more then welcome.

Paolo Abeni (3):
net: gro: avoid re-computing truesize twice on recycle
net: gro: minor optimization for dev_gro_receive()
net: gro: register gso and gro offload on separate lists

 include/linux/netdevice.h |   3 +-
 include/net/gro.h         |  13 +++--
 net/core/gro.c            | 107 +++++++++++++++++++++-----------------
 3 files changed, 70 insertions(+), 53 deletions(-)
Signature · 1 line
2.34.1
author maintainer
#2 lore raw

RFC PATCH 2/3 net: gro: minor optimization for dev_gro_receive()

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +9 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
+	u32	zeroed_end[0];
+
 	/* used to support CHECKSUM_COMPLETE for tunneling protocols */
 	__wsum	csum;
 
net/core/gro.c +5 -11
index d43d42215bdb..b9ebe9298731 100644
@@ -435,6 +435,9 @@ static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 	napi_gro_complete(napi, oldest);
 }
 
+#define zeroed_len	(offsetof(struct napi_gro_cb, zeroed_end) - \
+			 offsetof(struct napi_gro_cb, zeroed_start))
+
 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 {
 	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
@@ -459,29 +462,20 @@ static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff
 
 		skb_set_network_header(skb, skb_gro_offset(skb));
 		skb_reset_mac_len(skb);
-		NAPI_GRO_CB(skb)->same_flow = 0;
+		BUILD_BUG_ON(zeroed_len != sizeof(NAPI_GRO_CB(skb)->zeroed_start[0]));
+		NAPI_GRO_CB(skb)->zeroed_start[0] = 0;
 		NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
-		NAPI_GRO_CB(skb)->free = 0;
-		NAPI_GRO_CB(skb)->encap_mark = 0;
-		NAPI_GRO_CB(skb)->recursion_counter = 0;
-		NAPI_GRO_CB(skb)->is_fou = 0;
 		NAPI_GRO_CB(skb)->is_atomic = 1;
-		NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
 
 		/* Setup for GRO checksum validation */
 		switch (skb->ip_summed) {
 		case CHECKSUM_COMPLETE:
 			NAPI_GRO_CB(skb)->csum = skb->csum;
 			NAPI_GRO_CB(skb)->csum_valid = 1;
-			NAPI_GRO_CB(skb)->csum_cnt = 0;
 			break;
 		case CHECKSUM_UNNECESSARY:
 			NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
-			NAPI_GRO_CB(skb)->csum_valid = 0;
 			break;
-		default:
-			NAPI_GRO_CB(skb)->csum_cnt = 0;
-			NAPI_GRO_CB(skb)->csum_valid = 0;
 		}
 
 		pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
Signature · 1 line
2.34.1
author maintainer
#1 lore raw

RFC PATCH 1/3 net: gro: avoid re-computing truesize twice on recycle

In reply to Paolo Abeni

After commit 5e10da5385d2 ("skbuff: allow 'slow_gro' for skb carring sock reference") and commit af352460b465 ("net: fix GRO skb truesize update") the truesize of freed skb is properly updated by the GRO engine, we don't need anymore resetting it at recycle time.

---
 net/core/gro.c | 1 -
 1 file changed, 1 deletion(-)
net/core/gro.c +0 -1
index a11b286d1495..d43d42215bdb 100644
@@ -634,7 +634,6 @@ static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
 
 	skb->encapsulation = 0;
 	skb_shinfo(skb)->gso_type = 0;
-	skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 	if (unlikely(skb->slow_gro)) {
 		skb_orphan(skb);
 		skb_ext_reset(skb);
Signature · 1 line
2.34.1
author maintainer
#9 lore raw

RFC PATCH 3/3 net: gro: register gso and gro offload on separate lists

In reply to Paolo Abeni

So that we know each element in gro_list has valid gro callbacks (and the same for gso). This allows dropping a bunch of conditional in fastpath.

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

After:
0000000000000bb0 l F .text 0000000000000325 dev_gro_receive

---
 include/linux/netdevice.h |  3 +-
 net/core/gro.c            | 90 +++++++++++++++++++++++----------------
 2 files changed, 56 insertions(+), 37 deletions(-)
include/linux/netdevice.h +2 -1
index 3213c7227b59..406cb457d788 100644
@@ -2564,7 +2564,8 @@ struct packet_offload {
 	__be16			 type;	/* This is really htons(ether_type). */
 	u16			 priority;
 	struct offload_callbacks callbacks;
-	struct list_head	 list;
+	struct list_head	 gro_list;
+	struct list_head	 gso_list;
 };
 
 /* often modified stats are per-CPU, other are shared (netdev->stats) */
net/core/gro.c +54 -36

176 lines not on this page · raw message

Signature · 1 line
2.34.1
#3 lore raw

Re: [RFC PATCH 2/3] net: gro: minor optimization for dev_gro_receive()

In reply to Paolo Abeni

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 16:24:19 +0100

Quoted 50 lines

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +7 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;

include/net/gro.h struct napi_gro_cb { · line 76

Quoted 4 lines
include/net/gro.h +1 -0
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
+	u32	zeroed_end[0];

This should be wrapped in struct_group() I believe, or compilers will start complaining soon. See [0] for the details. Adding Kees to the CCs.

Quoted 52 lines
include/net/gro.h +1 -0
+
 	/* used to support CHECKSUM_COMPLETE for tunneling protocols */
 	__wsum	csum;
 
net/core/gro.c +5 -11
index d43d42215bdb..b9ebe9298731 100644
@@ -435,6 +435,9 @@ static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 	napi_gro_complete(napi, oldest);
 }
 
+#define zeroed_len	(offsetof(struct napi_gro_cb, zeroed_end) - \
+			 offsetof(struct napi_gro_cb, zeroed_start))
+
 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 {
 	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
@@ -459,29 +462,20 @@ static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff
 
 		skb_set_network_header(skb, skb_gro_offset(skb));
 		skb_reset_mac_len(skb);
-		NAPI_GRO_CB(skb)->same_flow = 0;
+		BUILD_BUG_ON(zeroed_len != sizeof(NAPI_GRO_CB(skb)->zeroed_start[0]));
+		NAPI_GRO_CB(skb)->zeroed_start[0] = 0;
 		NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
-		NAPI_GRO_CB(skb)->free = 0;
-		NAPI_GRO_CB(skb)->encap_mark = 0;
-		NAPI_GRO_CB(skb)->recursion_counter = 0;
-		NAPI_GRO_CB(skb)->is_fou = 0;
 		NAPI_GRO_CB(skb)->is_atomic = 1;
-		NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
 
 		/* Setup for GRO checksum validation */
 		switch (skb->ip_summed) {
 		case CHECKSUM_COMPLETE:
 			NAPI_GRO_CB(skb)->csum = skb->csum;
 			NAPI_GRO_CB(skb)->csum_valid = 1;
-			NAPI_GRO_CB(skb)->csum_cnt = 0;
 			break;
 		case CHECKSUM_UNNECESSARY:
 			NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
-			NAPI_GRO_CB(skb)->csum_valid = 0;
 			break;
-		default:
-			NAPI_GRO_CB(skb)->csum_cnt = 0;
-			NAPI_GRO_CB(skb)->csum_valid = 0;
 		}
 
 		pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
Signature · 1 line
2.34.1

[0] https://lore.kernel.org/linux-hardening/[email protected]

Thanks,
Al

author maintainer
#4 lore raw

Re: [RFC PATCH 2/3] net: gro: minor optimization for dev_gro_receive()

Quoted 56 lines · Alexander Lobakin

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 16:24:19 +0100

Quoted 53 lines

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +7 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
Quoted 5 lines

include/net/gro.h struct napi_gro_cb { · line 76

Quoted 1 line
include/net/gro.h +1 -0
+	u32	zeroed_end[0];

This should be wrapped in struct_group() I believe, or compilers will start complaining soon. See [0] for the details. Adding Kees to the CCs.

Thank you for the reference. That really slipped-off my mind.

This patch does not use memcpy() or similar, just a single direct assignement. Would that still require struct_group()?

Thanks!

Paolo

#5 lore raw

Re: [RFC PATCH 2/3] net: gro: minor optimization for dev_gro_receive()

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 17:31:00 +0100

Quoted 63 lines · Alexander Lobakin
Quoted 56 lines

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 16:24:19 +0100

Quoted 53 lines

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +7 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
Quoted 5 lines

include/net/gro.h struct napi_gro_cb { · line 76

Quoted 1 line
include/net/gro.h +1 -0
+	u32	zeroed_end[0];

This should be wrapped in struct_group() I believe, or compilers will start complaining soon. See [0] for the details. Adding Kees to the CCs.

Quoted 4 lines

Thank you for the reference. That really slipped-off my mind.

This patch does not use memcpy() or similar, just a single direct assignement. Would that still require struct_group()?

Oof, sorry, I saw start/end and overlooked that it's only for a single assignment.
Then it shouldn't cause warnings, but maybe use an anonymous union instead?

	union {
		u32 zeroed;
		struct {
			u16 gro_remcsum_start;
			...
		};
	};
	__wsum csum;

Use can still use a BUILD_BUG_ON() in this case, like sizeof(zeroed) != offsetof(csum) - offsetof(zeroed).

Quoted 4 lines

Thanks!

Paolo

Thanks,
Al

author maintainer
#6 lore raw

Re: [RFC PATCH 2/3] net: gro: minor optimization for dev_gro_receive()

Hello,

Quoted 85 lines · Alexander Lobakin

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 17:31:00 +0100

Quoted 63 lines · Alexander Lobakin
Quoted 56 lines

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 16:24:19 +0100

Quoted 53 lines

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +7 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
Quoted 5 lines

include/net/gro.h struct napi_gro_cb { · line 76

Quoted 1 line
include/net/gro.h +1 -0
+	u32	zeroed_end[0];

This should be wrapped in struct_group() I believe, or compilers will start complaining soon. See [0] for the details. Adding Kees to the CCs.

Quoted 4 lines

Thank you for the reference. That really slipped-off my mind.

This patch does not use memcpy() or similar, just a single direct assignement. Would that still require struct_group()?

Oof, sorry, I saw start/end and overlooked that it's only for a single assignment.
Then it shouldn't cause warnings, but maybe use an anonymous union instead?

	union {
		u32 zeroed;
		struct {
			u16 gro_remcsum_start;
			...
		};
	};
	__wsum csum;
Quoted 2 lines

Use can still use a BUILD_BUG_ON() in this case, like sizeof(zeroed) != offsetof(csum) - offsetof(zeroed).

Please forgive me for the very long delay. I'm looking again at this stuff for formal non-rfc submission.

I like the anonymous union less, because it will move around much more code - making the patch less readable - and will be more fragile e.g. some comment alike "please don't move around 'csum'" would be needed.

No strong opinion anyway, so if you really prefer that way I can adapt. Please let me know.

Thanks!

Paolo

#7 lore raw

Re: [RFC PATCH 2/3] net: gro: minor optimization for dev_gro_receive()

From: Paolo Abeni <[email protected]>
Date: Wed, 02 Feb 2022 11:09:41 +0100

Quoted 1 line

Hello,

Hi!

Quoted 87 lines · Alexander Lobakin
Quoted 85 lines

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 17:31:00 +0100

Quoted 63 lines · Alexander Lobakin
Quoted 56 lines

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 16:24:19 +0100

Quoted 53 lines

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +7 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
Quoted 5 lines

include/net/gro.h struct napi_gro_cb { · line 76

Quoted 1 line
include/net/gro.h +1 -0
+	u32	zeroed_end[0];

This should be wrapped in struct_group() I believe, or compilers will start complaining soon. See [0] for the details. Adding Kees to the CCs.

Quoted 4 lines

Thank you for the reference. That really slipped-off my mind.

This patch does not use memcpy() or similar, just a single direct assignement. Would that still require struct_group()?

Oof, sorry, I saw start/end and overlooked that it's only for a single assignment.
Then it shouldn't cause warnings, but maybe use an anonymous union instead?

	union {
		u32 zeroed;
		struct {
			u16 gro_remcsum_start;
			...
		};
	};
	__wsum csum;
Quoted 5 lines
Quoted 2 lines

Use can still use a BUILD_BUG_ON() in this case, like sizeof(zeroed) != offsetof(csum) - offsetof(zeroed).

Please forgive me for the very long delay. I'm looking again at this stuff for formal non-rfc submission.

Sure, not a problem at all (:

Quoted 4 lines

I like the anonymous union less, because it will move around much more code - making the patch less readable - and will be more fragile e.g. some comment alike "please don't move around 'csum'" would be needed.

We still need comments around zeroed_{start,end}[0] for now. I used offsetof(csum) as offsetofend(is_flist) which I'd prefer here unfortunately expands to offsetof(is_flist) + sizeof(is_flist), and the latter causes an error of using sizeof() against a bitfield.

Quoted 3 lines

No strong opinion anyway, so if you really prefer that way I can adapt. Please let me know.

I don't really have a strong preference here, I just suspect that zero-length array will produce or already produce -Warray-bounds warnings, and empty-struct constructs like

	struct { } zeroed_start;
	u16 gro_remcsum_start;
	...
	struct { } zeroed_end;

	memset(NAPI_GRO_CB(skb)->zeroed_start, 0,
	       offsetofend(zeroed_end) - offsetsof(zeroed_start));

will trigger Fortify compile-time errors from Kees' KSPP tree.

I think we could use

	__struct_group(/* no tag */, zeroed, /* no attrs */,
		u16 gro_remcsum_start;
		...
		u8 is_flist:1;
	);
	__wsum csum;

	BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
	BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
				 sizeof(u32))); /* Avoid slow unaligned acc */

	*(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;

This doesn't depend on csum, doesn't need struct { } or struct zero[0] markers and still uses a direct assignment.

Also adding Gustavo, maybe he'd like to leave a comment here.

Quoted 4 lines

Thanks!

Paolo

Thanks,
Al

#8 lore raw

Re: [RFC PATCH 2/3] net: gro: minor optimization for dev_gro_receive()

Quoted 74 lines · Alexander Lobakin

From: Paolo Abeni <[email protected]>
Date: Wed, 02 Feb 2022 11:09:41 +0100

Quoted 1 line

Hello,

Hi!

Quoted 67 lines · Alexander Lobakin
Quoted 65 lines

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 17:31:00 +0100

Quoted 62 lines · Alexander Lobakin
Quoted 61 lines

From: Paolo Abeni <[email protected]>
Date: Tue, 18 Jan 2022 16:24:19 +0100

Quoted 50 lines

While inspecting some perf report, I noticed that the compiler emits suboptimal code for the napi CB initialization, fetching and storing multiple times the memory for flags bitfield. This is with gcc 10.3.1, but I observed the same with older compiler versions.

We can help the compiler to do a nicer work e.g. initially setting all the bitfield to 0 using an u16 alias. The generated code is quite smaller, with the same number of conditional

Before:
objdump -t net/core/gro.o | grep " F .text"
0000000000000bb0 l F .text 0000000000000357 dev_gro_receive

After:
0000000000000bb0 l F .text 000000000000033c dev_gro_receive

---
 include/net/gro.h | 13 +++++++++----
 net/core/gro.c    | 16 +++++-----------
 2 files changed, 14 insertions(+), 15 deletions(-)
include/net/gro.h +7 -4
index 8f75802d50fd..a068b27d341f 100644
@@ -29,14 +29,17 @@ struct napi_gro_cb {
 	/* Number of segments aggregated. */
 	u16	count;
 
-	/* Start offset for remote checksum offload */
-	u16	gro_remcsum_start;
+	/* Used in ipv6_gro_receive() and foo-over-udp */
+	u16	proto;
 
 	/* jiffies when first packet was created/queued */
 	unsigned long age;
 
-	/* Used in ipv6_gro_receive() and foo-over-udp */
-	u16	proto;
+	/* portion of the cb set to zero at every gro iteration */
+	u32	zeroed_start[0];
+
+	/* Start offset for remote checksum offload */
+	u16	gro_remcsum_start;
 
 	/* This is non-zero if the packet may be of the same flow. */
 	u8	same_flow:1;

include/net/gro.h struct napi_gro_cb { · line 76

Quoted 4 lines
include/net/gro.h +1 -0
@@ -70,6 +73,8 @@ struct napi_gro_cb {
 	/* GRO is done by frag_list pointer chaining. */
 	u8	is_flist:1;
 
+	u32	zeroed_end[0];

This should be wrapped in struct_group() I believe, or compilers will start complaining soon. See [0] for the details. Adding Kees to the CCs.

Hi! Sorry I missed the original email sent my way. :)

Quoted 39 lines
Quoted 20 lines
Quoted 20 lines
Quoted 5 lines

Thank you for the reference. That really slipped-off my mind.

This patch does not use memcpy() or similar, just a single direct assignement. Would that still require struct_group()?

Oof, sorry, I saw start/end and overlooked that it's only for a single assignment.
Then it shouldn't cause warnings, but maybe use an anonymous union instead?

	union {
		u32 zeroed;
		struct {
			u16 gro_remcsum_start;
			...
		};
	};
	__wsum csum;
Quoted 5 lines
Quoted 2 lines

Use can still use a BUILD_BUG_ON() in this case, like sizeof(zeroed) != offsetof(csum) - offsetof(zeroed).

Please forgive me for the very long delay. I'm looking again at this stuff for formal non-rfc submission.

Sure, not a problem at all (:

Quoted 4 lines

I like the anonymous union less, because it will move around much more code - making the patch less readable - and will be more fragile e.g. some comment alike "please don't move around 'csum'" would be needed.

We still need comments around zeroed_{start,end}[0] for now. I used offsetof(csum) as offsetofend(is_flist) which I'd prefer here unfortunately expands to offsetof(is_flist) + sizeof(is_flist), and the latter causes an error of using sizeof() against a bitfield.

Quoted 1 line
Quoted 6 lines
Quoted 2 lines

No strong opinion anyway, so if you really prefer that way I can adapt. Please let me know.

I don't really have a strong preference here, I just suspect that zero-length array will produce or already produce -Warray-bounds warnings, and empty-struct constructs like

Yes, -Warray-bounds would yell very loudly about an 0-element array having its first element assigned. :)

Quoted 9 lines
	struct { } zeroed_start;
	u16 gro_remcsum_start;
	...
	struct { } zeroed_end;

	memset(NAPI_GRO_CB(skb)->zeroed_start, 0,
	       offsetofend(zeroed_end) - offsetsof(zeroed_start));
Quoted 5 lines

will trigger Fortify compile-time errors from Kees' KSPP tree.

I think we could use

	__struct_group(/* no tag */, zeroed, /* no attrs */,

Since this isn't UAPI, you can just do:

struct_group(zeroed,

Quoted 12 lines
		u16 gro_remcsum_start;
		...
		u8 is_flist:1;
	);
	__wsum csum;

	BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
	BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
				 sizeof(u32))); /* Avoid slow unaligned acc */

	*(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
Quoted 2 lines

This doesn't depend on csum, doesn't need struct { } or struct zero[0] markers and still uses a direct assignment.

Given the kernel's -O2 use, a constant-sized u32 memcpy() and a direct assignment will resolve to the same machine code:

https://godbolt.org/z/b9qKexx6q

void zero_var(struct object *p)

{
    p->zero_var = 0;
}

void zero_group(struct object *p)

{
    memset(&p->zero_group, 0, sizeof(p->zero_group));
}

zero_var:
movl $0, (%rdi)
ret
zero_group:
movl $0, (%rdi)
ret

And assigning them individually results in careful "and"ing to avoid the padding:

zero_each:
andl $-134217728, (%rdi)
ret

I would have recommended struct_group() + memset() here, just because you don't need the casts. But since you're doing BUILD_BUG_ON() size verification, it's fine either way.

If you want to avoid the cast and avoid the memset(), you could use the earlier suggestion of anonymous union and anonymous struct. It's basically an open-coded struct_group, but you get to pick the type of the overlapping variable. :)

-Kees

Quoted 9 lines

Also adding Gustavo, maybe he'd like to leave a comment here.

Quoted 4 lines

Thanks!

Paolo

Thanks,
Al

Signature · 1 line
Kees Cook