numam-dpdk/drivers/net/sfc/sfc_tweak.h
Igor Romanov 96fd2bd69b net/sfc: support flow action count in transfer rules
For now, a rule may have only one dedicated counter, shared counters
are not supported.

HW delivers (or "streams") counter readings using special packets.
The driver creates a dedicated Rx queue to receive such packets
and requests that HW start "streaming" the readings to it.

The counter queue is polled periodically, and the first available
service core is used for that. Hence, the user has to specify at least
one service core for counters to work. Such a core is shared by all
MAE-capable devices managed by sfc driver.

Signed-off-by: Igor Romanov <igor.romanov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Reviewed-by: Andy Moreton <amoreton@xilinx.com>
Reviewed-by: Ivan Malov <ivan.malov@oktetlabs.ru>
2021-07-20 12:20:31 +02:00

55 lines
1.8 KiB
C

/* SPDX-License-Identifier: BSD-3-Clause
*
* Copyright(c) 2019-2021 Xilinx, Inc.
* Copyright(c) 2016-2019 Solarflare Communications Inc.
*
* This software was jointly developed between OKTET Labs (under contract
* for Solarflare) and Solarflare Communications, Inc.
*/
#ifndef _SFC_TWEAK_H_
#define _SFC_TWEAK_H_
/*
* The header is intended to collect defines/constants which could be
* tweaked to improve the PMD performance characteristics depending on
* the usecase or requirements (CPU load, packet rate, latency).
*/
/**
* Number of Rx descriptors in the bulk submitted on Rx ring refill.
*/
#define SFC_RX_REFILL_BULK (RTE_CACHE_LINE_SIZE / sizeof(efx_qword_t))
/**
* Make the transmit path reap at least one time per a burst;
* this improves cache locality because the same mbufs may be used to send
* subsequent bursts in certain cases because of well-timed reap
*/
#define SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE 1
/** Default free threshold follows recommendations from DPDK documentation */
#define SFC_TX_DEFAULT_FREE_THRESH 32
/** Number of mbufs to be freed in bulk in a single call */
#define SFC_TX_REAP_BULK_SIZE 32
/**
* Default head-of-line block timeout to wait for Rx descriptor before
* packet drop because of no descriptors available.
*
* DPDK FW variant only with equal stride super-buffer Rx mode.
*/
#define SFC_RXD_WAIT_TIMEOUT_NS_DEF (200U * 1000)
/**
* Ideally reading packet and byte counters together should return
* consistent values. I.e. a number of bytes corresponds to a number of
* packets. Since counters are updated in one thread and queried in
* another it requires either locking or atomics which are very
* expensive from performance point of view. So, disable it by default.
*/
#define SFC_SW_STATS_ATOMIC 0
#endif /* _SFC_TWEAK_H_ */