This change re-arranges the fields within the tcp-pcb so that

they are more in order of cache line use as one passes
through the tcp_input/output paths (non-errors most likely path). This
helps speed up cache line optimization so that the tcp stack runs
a bit more efficently.

Sponsored by:	Netflix Inc.
Differential Revision:	https://reviews.freebsd.org/D15136
This commit is contained in:
Randall Stewart 2018-04-26 21:41:16 +00:00
parent d40a1b91c0
commit 803a230592
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/head/; revision=333041

View File

@ -83,125 +83,123 @@ STAILQ_HEAD(tcp_log_stailq, tcp_log_mem);
/*
* Tcp control block, one per tcp; fields:
* Organized for 16 byte cacheline efficiency.
* Organized for 64 byte cacheline efficiency based
* on common tcp_input/tcp_output processing.
*/
struct tcpcb {
struct tsegqe_head t_segq; /* segment reassembly queue */
int t_segqlen; /* segment reassembly queue length */
int t_dupacks; /* consecutive dup acks recd */
struct mbuf *t_in_pkt; /* head of the input packet queue for the tcp_hpts system */
struct mbuf *t_tail_pkt; /* tail of the input packet queue for the tcp_hpts system */
struct tcp_timer *t_timers; /* All the TCP timers in one struct */
/* Cache line 1 */
struct inpcb *t_inpcb; /* back pointer to internet pcb */
int t_state; /* state of this connection */
struct tcp_function_block *t_fb;/* TCP function call block */
void *t_fb_ptr; /* Pointer to t_fb specific data */
uint32_t t_maxseg:24, /* maximum segment size */
t_logstate:8; /* State of "black box" logging */
uint32_t t_state:4, /* state of this connection */
bits_spare : 24;
u_int t_flags;
struct vnet *t_vnet; /* back pointer to parent vnet */
tcp_seq snd_una; /* sent but unacknowledged */
tcp_seq snd_max; /* highest sequence number sent;
* used to recognize retransmits
*/
tcp_seq snd_nxt; /* send next */
tcp_seq snd_up; /* send urgent pointer */
tcp_seq snd_wl1; /* window update seg seq number */
tcp_seq snd_wl2; /* window update seg ack number */
tcp_seq iss; /* initial send sequence number */
tcp_seq irs; /* initial receive sequence number */
uint32_t snd_wnd; /* send window */
uint32_t snd_cwnd; /* congestion-controlled window */
uint32_t cl1_spare; /* Spare to round out CL 1 */
/* Cache line 2 */
u_int32_t ts_offset; /* our timestamp offset */
u_int32_t rfbuf_ts; /* recv buffer autoscaling timestamp */
int rcv_numsacks; /* # distinct sack blks present */
u_int t_tsomax; /* TSO total burst length limit in bytes */
u_int t_tsomaxsegcount; /* TSO maximum segment count */
u_int t_tsomaxsegsize; /* TSO maximum segment size in bytes */
tcp_seq rcv_nxt; /* receive next */
tcp_seq rcv_adv; /* advertised window */
uint32_t rcv_wnd; /* receive window */
u_int t_flags2; /* More tcpcb flags storage */
int t_srtt; /* smoothed round-trip time */
int t_rttvar; /* variance in round-trip time */
u_int32_t ts_recent; /* timestamp echo data */
u_char snd_scale; /* window scaling for send window */
u_char rcv_scale; /* window scaling for recv window */
u_char snd_limited; /* segments limited transmitted */
u_char request_r_scale; /* pending window scaling */
tcp_seq last_ack_sent;
u_int t_rcvtime; /* inactivity time */
/* Cache line 3 */
tcp_seq rcv_up; /* receive urgent pointer */
uint32_t snd_wnd; /* send window */
uint32_t snd_cwnd; /* congestion-controlled window */
int t_segqlen; /* segment reassembly queue length */
struct tsegqe_head t_segq; /* segment reassembly queue */
struct mbuf *t_in_pkt;
struct mbuf *t_tail_pkt;
struct tcp_timer *t_timers; /* All the TCP timers in one struct */
struct vnet *t_vnet; /* back pointer to parent vnet */
uint32_t snd_ssthresh; /* snd_cwnd size threshold for
* for slow start exponential to
* linear switch
*/
tcp_seq snd_wl1; /* window update seg seq number */
/* Cache line 4 */
tcp_seq snd_wl2; /* window update seg ack number */
tcp_seq irs; /* initial receive sequence number */
tcp_seq iss; /* initial send sequence number */
u_int t_acktime;
u_int ts_recent_age; /* when last updated */
tcp_seq snd_recover; /* for use in NewReno Fast Recovery */
u_int t_rcvtime; /* inactivity time */
u_int t_starttime; /* time connection was established */
u_int t_rtttime; /* RTT measurement start time */
tcp_seq t_rtseq; /* sequence number being timed */
int t_rxtcur; /* current retransmit value (ticks) */
u_int t_maxseg; /* maximum segment size */
u_int t_pmtud_saved_maxseg; /* pre-blackhole MSS */
int t_srtt; /* smoothed round-trip time */
int t_rttvar; /* variance in round-trip time */
int t_rxtshift; /* log(2) of rexmt exp. backoff */
u_int t_rttmin; /* minimum rtt allowed */
u_int t_rttbest; /* best rtt we've seen */
u_long t_rttupdated; /* number of times rtt sampled */
uint32_t max_sndwnd; /* largest window peer has offered */
int t_softerror; /* possible error not yet reported */
/* out-of-band data */
uint16_t cl4_spare; /* Spare to adjust CL 4 */
char t_oobflags; /* have some */
char t_iobc; /* input character */
/* RFC 1323 variables */
u_char snd_scale; /* window scaling for send window */
u_char rcv_scale; /* window scaling for recv window */
u_char request_r_scale; /* pending window scaling */
u_int32_t ts_recent; /* timestamp echo data */
u_int ts_recent_age; /* when last updated */
u_int32_t ts_offset; /* our timestamp offset */
int t_rxtcur; /* current retransmit value (ticks) */
tcp_seq last_ack_sent;
/* experimental */
int t_rxtshift; /* log(2) of rexmt exp. backoff */
u_int t_rtttime; /* RTT measurement start time */
tcp_seq t_rtseq; /* sequence number being timed */
u_int t_starttime; /* time connection was established */
u_int t_pmtud_saved_maxseg; /* pre-blackhole MSS */
u_int t_rttmin; /* minimum rtt allowed */
u_int t_rttbest; /* best rtt we've seen */
int t_softerror; /* possible error not yet reported */
uint32_t max_sndwnd; /* largest window peer has offered */
/* Cache line 5 */
uint32_t snd_cwnd_prev; /* cwnd prior to retransmit */
uint32_t snd_ssthresh_prev; /* ssthresh prior to retransmit */
tcp_seq snd_recover_prev; /* snd_recover prior to retransmit */
int t_sndzerowin; /* zero-window updates sent */
u_int t_badrxtwin; /* window for retransmit recovery */
u_char snd_limited; /* segments limited transmitted */
/* SACK related state */
u_long t_rttupdated; /* number of times rtt sampled */
int snd_numholes; /* number of holes seen by sender */
u_int t_badrxtwin; /* window for retransmit recovery */
TAILQ_HEAD(sackhole_head, sackhole) snd_holes;
/* SACK scoreboard (sorted) */
tcp_seq snd_fack; /* last seq number(+1) sack'd by rcv'r*/
int rcv_numsacks; /* # distinct sack blks present */
struct sackblk sackblks[MAX_SACK_BLKS]; /* seq nos. of sack blocks */
tcp_seq sack_newdata; /* New data xmitted in this recovery
episode starts at this seq number */
struct sackblk sackblks[MAX_SACK_BLKS]; /* seq nos. of sack blocks */
struct sackhint sackhint; /* SACK scoreboard hint */
int t_rttlow; /* smallest observerved RTT */
u_int32_t rfbuf_ts; /* recv buffer autoscaling timestamp */
int rfbuf_cnt; /* recv buffer autoscaling byte count */
struct toedev *tod; /* toedev handling this connection */
int t_sndrexmitpack; /* retransmit packets sent */
int t_rcvoopack; /* out-of-order packets received */
void *t_toe; /* TOE pcb pointer */
int t_bytes_acked; /* # bytes acked during current RTT */
struct cc_algo *cc_algo; /* congestion control algorithm */
struct cc_var *ccv; /* congestion control specific vars */
struct osd *osd; /* storage for Khelp module data */
int t_bytes_acked; /* # bytes acked during current RTT */
u_int t_keepinit; /* time to establish connection */
u_int t_keepidle; /* time before keepalive probes begin */
u_int t_keepintvl; /* interval between keepalives */
u_int t_keepcnt; /* number of keepalives before close */
u_int t_tsomax; /* TSO total burst length limit in bytes */
u_int t_tsomaxsegcount; /* TSO maximum segment count */
u_int t_tsomaxsegsize; /* TSO maximum segment size in bytes */
u_int t_flags2; /* More tcpcb flags storage */
int t_logstate; /* State of "black box" logging */
struct tcp_log_stailq t_logs; /* Log buffer */
int t_dupacks; /* consecutive dup acks recd */
int t_lognum; /* Number of log entries */
uint32_t t_logsn; /* Log "serial number" */
struct tcp_log_stailq t_logs; /* Log buffer */
struct tcp_log_id_node *t_lin;
struct tcp_log_id_bucket *t_lib;
const char *t_output_caller; /* Function that called tcp_output */
struct tcp_function_block *t_fb;/* TCP function call block */
void *t_fb_ptr; /* Pointer to t_fb specific data */
uint32_t t_logsn; /* Log "serial number" */
uint8_t t_tfo_client_cookie_len; /* TCP Fast Open client cookie length */
unsigned int *t_tfo_pending; /* TCP Fast Open server pending counter */
union {