net: skbuff: hide some bitfield members
Subsystems (4)
Jakub Kicinski
maintainerv2 cover net: skbuff: hide some bitfield members
There is a number of protocol or subsystem specific fields in struct sk_buff which are only accessed by one subsystem. We can wrap them in ifdefs with minimal code impact.
This gives us a better chance to save a 2B and a 4B holes resulting with the following savings (assuming a lucky kernel config):
- /* size: 232, cachelines: 4, members: 28 */
- /* sum members: 227, holes: 1, sum holes: 4 */
- /* sum bitfield members: 8 bits (1 bytes) */
+ /* size: 224, cachelines: 4, members: 28 */
/* forced alignments: 2 */
- /* last cacheline: 40 bytes */
+ /* last cacheline: 32 bytes */
I think that the changes shouldn't be too controversial. The only one I'm not 100% sure of is the SCTP one, 12 extra LoC for one bit.. But it did fit squarely in the "this bit has only one user" category.
v2:
- use IS_ENABLED() for nf_trace
v1: https://lore.kernel.org/all/[email protected]/
Jakub Kicinski (5):
net: skbuff: hide wifi_acked when CONFIG_WIRELESS not set
net: skbuff: hide csum_not_inet when CONFIG_IP_SCTP not set
net: skbuff: move alloc_cpu into a potential hole
net: skbuff: push nf_trace down the bitfield
net: skbuff: hide nf_trace and ipvs_property
include/linux/skbuff.h | 38 ++++++++++++++++++++++++++++++++++---- include/net/sock.h | 2 +- net/core/dev.c | 3 +-- net/core/skbuff.c | 2 ++ net/sched/act_csum.c | 3 +-- net/socket.c | 2 ++ 6 files changed, 41 insertions(+), 9 deletions(-)
Signature · 1 line
2.39.2
Jakub Kicinski
maintainerPATCH v2 1/5 net: skbuff: hide wifi_acked when CONFIG_WIRELESS not set
Datacenter kernel builds will very likely not include WIRELESS, so let them shave 2 bits off the skb by hiding the wifi fields.
---
CC: [email protected]
--- include/linux/skbuff.h | 11 +++++++++++ include/net/sock.h | 2 +- net/core/skbuff.c | 2 ++ net/socket.c | 2 ++ 4 files changed, 16 insertions(+), 1 deletion(-)
include/linux/skbuff.h +11 -0
@@ -953,8 +953,10 @@ struct sk_buff { __u8 l4_hash:1; __u8 sw_hash:1; +#ifdef CONFIG_WIRELESS __u8 wifi_acked_valid:1; __u8 wifi_acked:1; +#endif __u8 no_fcs:1;/* Indicates the inner headers are valid in the skbuff. */ __u8 encapsulation:1;
@@ -1187,6 +1189,15 @@ static inline unsigned int skb_napi_id(const struct sk_buff *skb)#endif } +static inline bool skb_wifi_acked_valid (const struct sk_buff *skb) +{ +#ifdef CONFIG_WIRELESS +return skb->wifi_acked_valid; +#else +return 0 ; +#endif +} +/** * skb_unref - decrement the skb's reference count * @skb: buffer
include/net/sock.h +1 -1
@@ -2697,7 +2697,7 @@ sock_recv_timestamp(struct msghdr *msg, struct sock *sk, struct sk_buff *skb)else sock_write_timestamp (sk, kt); -if (sock_flag (sk,SOCK_WIFI_STATUS ) && skb->wifi_acked_valid) +if (sock_flag (sk,SOCK_WIFI_STATUS ) &&skb_wifi_acked_valid (skb))__sock_recv_wifi_status (msg, sk, skb); }
net/core/skbuff.c +2 -0
@@ -5189,6 +5189,7 @@ void skb_tstamp_tx(struct sk_buff *orig_skb, }EXPORT_SYMBOL_GPL (skb_tstamp_tx); +#ifdef CONFIG_WIRELESS void skb_complete_wifi_ack (struct sk_buff *skb,bool acked) {struct sock *sk = skb->sk;
@@ -5214,6 +5215,7 @@ void skb_complete_wifi_ack(struct sk_buff *skb, bool acked) kfree_skb(skb); }EXPORT_SYMBOL_GPL (skb_complete_wifi_ack); +#endif /* CONFIG_WIRELESS */ /** * skb_partial_csum_set - set up and verify partial csum values for packet
net/socket.c +2 -0
@@ -957,6 +957,7 @@ void __sock_recv_timestamp(struct msghdr *msg, struct sock *sk, }EXPORT_SYMBOL_GPL (__sock_recv_timestamp); +#ifdef CONFIG_WIRELESS void __sock_recv_wifi_status (struct msghdr *msg,struct sock *sk,struct sk_buff *skb) {
@@ -972,6 +973,7 @@ void __sock_recv_wifi_status(struct msghdr *msg, struct sock *sk, put_cmsg(msg, SOL_SOCKET, SCM_WIFI_STATUS, sizeof(ack), &ack); }EXPORT_SYMBOL_GPL (__sock_recv_wifi_status); +#endif static inline void sock_recv_drops (struct msghdr *msg,struct sock *sk,struct sk_buff *skb)
Signature · 1 line
2.39.2
Re: [PATCH net-next v2 1/5] net: skbuff: hide wifi_acked when CONFIG_WIRELESS not set
Quoted 2 lines · Jakub Kicinski
Datacenter kernel builds will very likely not include WIRELESS, so let them shave 2 bits off the skb by hiding the wifi fields.
:-)
You could make this be CONFIG_MAC80211 if you wanted to, but I don't mind CONFIG_WIRELESS and honestly they're probably pretty much equivalent because there aren't many non-mac80211 drivers.
johannes
Jakub Kicinski
maintainerPATCH v2 2/5 net: skbuff: hide csum_not_inet when CONFIG_IP_SCTP not set
In reply to Jakub KicinskiSCTP is not universally deployed, allow hiding its bit from the skb.
--- include/linux/skbuff.h | 14 ++++++++++++++ net/core/dev.c | 3 +-- net/sched/act_csum.c | 3 +-- 3 files changed, 16 insertions(+), 4 deletions(-)
include/linux/skbuff.h +14 -0
@@ -983,7 +983,9 @@ struct sk_buff { __u8 decrypted:1;#endif __u8 slow_gro:1; +#if IS_ENABLED (CONFIG_IP_SCTP ) __u8 csum_not_inet:1; +#endif #ifdef CONFIG_NET_SCHED __u16 tc_index;/* traffic control index */
@@ -5060,7 +5062,19 @@ static inline void skb_reset_redirect(struct sk_buff *skb)static inline bool skb_csum_is_sctp (struct sk_buff *skb) { +#if IS_ENABLED (CONFIG_IP_SCTP )return skb->csum_not_inet; +#else +return 0 ; +#endif +} + +static inline void skb_reset_csum_not_inet (struct sk_buff *skb) +{ + skb->ip_summed =CHECKSUM_NONE ; +#if IS_ENABLED (CONFIG_IP_SCTP ) + skb->csum_not_inet =0 ; +#endif } static inline void skb_set_kcov_handle(struct sk_buff *skb,
net/core/dev.c +1 -2
@@ -3315,8 +3315,7 @@ int skb_crc32c_csum_help(struct sk_buff *skb) skb->len - start, ~(__u32 )0 , crc32c_csum_stub)); *(__le32 *)(skb->data + offset) = crc32c_csum; - skb->ip_summed =CHECKSUM_NONE ; - skb->csum_not_inet =0 ; +skb_reset_csum_not_inet (skb); out:return ret; }
net/sched/act_csum.c +1 -2
@@ -376,8 +376,7 @@ static int tcf_csum_sctp(struct sk_buff *skb, unsigned int ihl, sctph->checksum =sctp_compute_cksum (skb,skb_network_offset (skb) + ihl); - skb->ip_summed =CHECKSUM_NONE ; - skb->csum_not_inet =0 ; +skb_reset_csum_not_inet (skb);return 1 ; }
Signature · 1 line
2.39.2
Jakub Kicinski
maintainerPATCH v2 3/5 net: skbuff: move alloc_cpu into a potential hole
In reply to Jakub Kicinskialloc_cpu is currently between 4 byte fields, so it's almost guaranteed to create a 2B hole. It has a knock on effect of creating a 4B hole after @end (and @end and @tail being in different cachelines).
None of this matters hugely, but for kernel configs which don't enable all the features there may well be a 2B hole after the bitfield. Move alloc_cpu there.
--- include/linux/skbuff.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
include/linux/skbuff.h +2 -1
@@ -991,6 +991,8 @@ struct sk_buff {__u16 tc_index;/* traffic control index */ #endif +u16 alloc_cpu; + union { __wsum csum; struct {
@@ -1014,7 +1016,6 @@ struct sk_buff { unsigned int sender_cpu; };#endif -u16 alloc_cpu;#ifdef CONFIG_NETWORK_SECMARK __u32 secmark;#endif
Signature · 1 line
2.39.2
Jakub Kicinski
maintainerPATCH v2 4/5 net: skbuff: push nf_trace down the bitfield
In reply to Jakub Kicinskinf_trace is a debug feature, AFAIU, and yet it sits oddly high in the sk_buff bitfield. Move it down, pushing up dst_pending_confirm and inner_protocol_type.
Next change will make nf_trace optional (under Kconfig) and all optional fields should be placed after 2b fields to avoid 2b fields straddling bytes.
dst_pending_confirm is L3, so it makes sense next to ignore_df. inner_protocol_type goes up just to keep the balance.
---
CC: [email protected]
CC: [email protected]
--- include/linux/skbuff.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
include/linux/skbuff.h +3 -3
@@ -934,7 +934,7 @@ struct sk_buff {/* public: */ __u8 pkt_type:3; /* see PKT_TYPE_MAX */ __u8 ignore_df:1; - __u8 nf_trace:1; + __u8 dst_pending_confirm:1; __u8 ip_summed:2; __u8 ooo_okay:1;
@@ -949,7 +949,7 @@ struct sk_buff { __u8 remcsum_offload:1; __u8 csum_complete_sw:1; __u8 csum_level:2; - __u8 dst_pending_confirm:1; + __u8 inner_protocol_type:1; __u8 l4_hash:1; __u8 sw_hash:1;
@@ -967,7 +967,7 @@ struct sk_buff {#endif __u8 ipvs_property:1; - __u8 inner_protocol_type:1; + __u8 nf_trace:1;#ifdef CONFIG_NET_SWITCHDEV __u8 offload_fwd_mark:1; __u8 offload_l3_fwd_mark:1;
Signature · 1 line
2.39.2
Jakub Kicinski
maintainerPATCH v2 5/5 net: skbuff: hide nf_trace and ipvs_property
In reply to Jakub KicinskiAccesses to nf_trace and ipvs_property are already wrapped by ifdefs where necessary. Don't allocate the bits for those fields at all if possible.
---
v2:
- use IS_ENABLED()
v1: https://lore.kernel.org/all/[email protected]/
--- include/linux/skbuff.h | 4 ++++ 1 file changed, 4 insertions(+)
include/linux/skbuff.h +4 -0
@@ -966,8 +966,12 @@ struct sk_buff { __u8 ndisc_nodetype:2;#endif +#if IS_ENABLED (CONFIG_IP_VS ) __u8 ipvs_property:1; +#endif +#if IS_ENABLED (CONFIG_NETFILTER_XT_TARGET_TRACE ) ||IS_ENABLED (CONFIG_NF_TABLES ) __u8 nf_trace:1; +#endif #ifdef CONFIG_NET_SWITCHDEV __u8 offload_fwd_mark:1; __u8 offload_l3_fwd_mark:1;
Signature · 1 line
2.39.2
Quoted 16 lines · On Mon, 17 Apr 2023 08:53:45 -0700 Jakub Kicinski wrote:
There is a number of protocol or subsystem specific fields in struct
sk_buffwhich are only accessed by one subsystem. We can wrap them in ifdefs with minimal code impact.This gives us a better chance to save a 2B and a 4B holes resulting with the following savings (assuming a lucky kernel config):
(quoted diff) +2
-4- /* size: 232, cachelines: 4, members: 28 */ - /* sum members: 227, holes: 1, sum holes: 4 */ - /* sum bitfield members: 8 bits (1 bytes) */ + /* size: 224, cachelines: 4, members: 28 */ /* forced alignments: 2 */ - /* last cacheline: 40 bytes */ + /* last cacheline: 32 bytes */
Quoted 4 lines
I think that the changes shouldn't be too controversial. The only one I'm not 100% sure of is the SCTP one, 12 extra LoC for one bit.. But it did fit squarely in the "this bit has only one user" category.
Missed Simon's tag, sorry about that:
https://lore.kernel.org/all/[email protected]/
This series was applied to netdev/net-next.git (main) by David S. Miller <[email protected]>:
Hello:
This series was applied to netdev/net-next.git (main) by David S. Miller <[email protected]>:
Quoted 9 lines · On Mon, 17 Apr 2023 08:53:45 -0700 you wrote:
There is a number of protocol or subsystem specific fields in struct
sk_buffwhich are only accessed by one subsystem. We can wrap them in ifdefs with minimal code impact.This gives us a better chance to save a 2B and a 4B holes resulting with the following savings (assuming a lucky kernel config):
[...]
Here is the summary with links:
- [net-next,v2,1/5] net: skbuff: hide wifi_acked when CONFIG_WIRELESS not set
https://git.kernel.org/netdev/net-next/c/eb6fba7555a8 - [net-next,v2,2/5] net: skbuff: hide csum_not_inet when CONFIG_IP_SCTP not set https://git.kernel.org/netdev/net-next/c/c24831a13ba2 - [net-next,v2,3/5] net: skbuff: move alloc_cpu into a potential hole https://git.kernel.org/netdev/net-next/c/4398f3f6d138 - [net-next,v2,4/5] net: skbuff: push nf_trace down the bitfield https://git.kernel.org/netdev/net-next/c/4c60d04c2888 - [net-next,v2,5/5] net: skbuff: hide nf_trace and ipvs_property https://git.kernel.org/netdev/net-next/c/48d80c394d3d
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html