From 245d157651bd6e7e588b1e4a1ae25c1f7bbd30fa Mon Sep 17 00:00:00 2001 From: Andrew Rybchenko Date: Tue, 17 Mar 2015 08:14:46 +0000 Subject: [PATCH] sfxge: add tunables to control LRO parameters on driver load time Sponsored by: Solarflare Communications, Inc. Approved by: gnn (mentor) --- share/man/man4/sfxge.4 | 21 +++++++++++++++++++++ sys/dev/sfxge/sfxge_rx.c | 40 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 61 insertions(+) diff --git a/share/man/man4/sfxge.4 b/share/man/man4/sfxge.4 index 09b59eb8f401..c745368ead16 100644 --- a/share/man/man4/sfxge.4 +++ b/share/man/man4/sfxge.4 @@ -120,6 +120,27 @@ The value must be greater than or equal to 0. The maximum number of allocated RSS channels for the Nth adapter. If set to 0 or unset, the number of channels is determined by the number of CPU cores. +.It Va hw.sfxge.lro.table_size +Size of the LRO hash table. +Must be a power of 2. +A larger table means we can accelerate a larger number of streams. +.It Va hw.sfxge.lro.chain_max +The maximum length of a hash chain. +If chains get too long then the lookup time increases and may exceed +the benefit of LRO. +.It Va hw.sfxge.lro.idle_ticks +The maximum time (in ticks) that a connection can be idle before it's LRO +state is discarded. +.It Va hw.sfxge.lro.slow_start_packets +Number of packets with payload that must arrive in-order before a connection +is eligible for LRO. +The idea is we should avoid coalescing segments when the sender is in +slow-start because reducing the ACK rate can damage performance. +.It Va hw.sfxge.lro.loss_packets +Number of packets with payload that must arrive in-order following loss +before a connection is eligible for LRO. +The idea is we should avoid coalescing segments when the sender is recovering +from loss, because reducing the ACK rate can damage performance. .El .Sh SUPPORT For general information and support, diff --git a/sys/dev/sfxge/sfxge_rx.c b/sys/dev/sfxge/sfxge_rx.c index b0cbb8a3f5b6..dbec7d98bedf 100644 --- a/sys/dev/sfxge/sfxge_rx.c +++ b/sys/dev/sfxge/sfxge_rx.c @@ -36,6 +36,7 @@ __FBSDID("$FreeBSD$"); #include #include #include +#include #include #include @@ -58,20 +59,38 @@ __FBSDID("$FreeBSD$"); #ifdef SFXGE_LRO +SYSCTL_NODE(_hw_sfxge, OID_AUTO, lro, CTLFLAG_RD, NULL, + "Large receive offload (LRO) parameters"); + +#define SFXGE_LRO_PARAM(_param) SFXGE_PARAM(lro._param) + /* Size of the LRO hash table. Must be a power of 2. A larger table * means we can accelerate a larger number of streams. */ static unsigned lro_table_size = 128; +TUNABLE_INT(SFXGE_LRO_PARAM(table_size), &lro_table_size); +SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, table_size, CTLFLAG_RDTUN, + &lro_table_size, 0, + "Size of the LRO hash table (must be a power of 2)"); /* Maximum length of a hash chain. If chains get too long then the lookup * time increases and may exceed the benefit of LRO. */ static unsigned lro_chain_max = 20; +TUNABLE_INT(SFXGE_LRO_PARAM(chain_max), &lro_chain_max); +SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, chain_max, CTLFLAG_RDTUN, + &lro_chain_max, 0, + "The maximum length of a hash chain"); /* Maximum time (in ticks) that a connection can be idle before it's LRO * state is discarded. */ static unsigned lro_idle_ticks; /* initialised in sfxge_rx_init() */ +TUNABLE_INT(SFXGE_LRO_PARAM(idle_ticks), &lro_idle_ticks); +SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, idle_ticks, CTLFLAG_RDTUN, + &lro_idle_ticks, 0, + "The maximum time (in ticks) that a connection can be idle " + "before it's LRO state is discarded"); /* Number of packets with payload that must arrive in-order before a * connection is eligible for LRO. The idea is we should avoid coalescing @@ -79,6 +98,11 @@ static unsigned lro_idle_ticks; /* initialised in sfxge_rx_init() */ * can damage performance. */ static int lro_slow_start_packets = 2000; +TUNABLE_INT(SFXGE_LRO_PARAM(slow_start_packets), &lro_slow_start_packets); +SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, slow_start_packets, CTLFLAG_RDTUN, + &lro_slow_start_packets, 0, + "Number of packets with payload that must arrive in-order before " + "a connection is eligible for LRO"); /* Number of packets with payload that must arrive in-order following loss * before a connection is eligible for LRO. The idea is we should avoid @@ -86,6 +110,11 @@ static int lro_slow_start_packets = 2000; * reducing the ACK rate can damage performance. */ static int lro_loss_packets = 20; +TUNABLE_INT(SFXGE_LRO_PARAM(loss_packets), &lro_loss_packets); +SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, loss_packets, CTLFLAG_RDTUN, + &lro_loss_packets, 0, + "Number of packets with payload that must arrive in-order " + "following loss before a connection is eligible for LRO"); /* Flags for sfxge_lro_conn::l2_id; must not collide with EVL_VLID_MASK */ #define SFXGE_LRO_L2_ID_VLAN 0x4000 @@ -1239,6 +1268,13 @@ sfxge_rx_init(struct sfxge_softc *sc) int rc; #ifdef SFXGE_LRO + if (!ISP2(lro_table_size)) { + log(LOG_ERR, "%s=%u must be power of 2", + SFXGE_LRO_PARAM(table_size), lro_table_size); + rc = EINVAL; + goto fail_lro_table_size; + } + if (lro_idle_ticks == 0) lro_idle_ticks = hz / 10 + 1; /* 100 ms */ #endif @@ -1266,5 +1302,9 @@ sfxge_rx_init(struct sfxge_softc *sc) sfxge_rx_qfini(sc, index); sc->rxq_count = 0; + +#ifdef SFXGE_LRO +fail_lro_table_size: +#endif return (rc); }