numam-dpdk/drivers/net/sfc/sfc_ef10.h
Andrew Rybchenko 141d287067 net/sfc: support aarch64 architecture
Enable the PMD build on aarch64.

Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andy Moreton <amoreton@xilinx.com>
2020-11-03 23:35:06 +01:00

143 lines
3.6 KiB
C

/* SPDX-License-Identifier: BSD-3-Clause
*
* Copyright(c) 2019-2020 Xilinx, Inc.
* Copyright(c) 2017-2019 Solarflare Communications Inc.
*
* This software was jointly developed between OKTET Labs (under contract
* for Solarflare) and Solarflare Communications, Inc.
*/
#ifndef _SFC_EF10_H
#define _SFC_EF10_H
#include "sfc_debug.h"
#ifdef __cplusplus
extern "C" {
#endif
/* Number of events in one cache line */
#define SFC_EF10_EV_PER_CACHE_LINE \
(RTE_CACHE_LINE_SIZE / sizeof(efx_qword_t))
#define SFC_EF10_EV_QCLEAR_MASK (~(SFC_EF10_EV_PER_CACHE_LINE - 1))
/*
* Use simple libefx-based implementation of the
* sfc_ef10_ev_qclear_cache_line() if SSE2 is not available
* since optimized implementation uses __m128i intrinsics.
*/
#ifndef __SSE2__
#define SFC_EF10_EV_QCLEAR_USE_EFX
#endif
#if defined(SFC_EF10_EV_QCLEAR_USE_EFX)
static inline void
sfc_ef10_ev_qclear_cache_line(void *ptr)
{
efx_qword_t *entry = ptr;
unsigned int i;
for (i = 0; i < SFC_EF10_EV_PER_CACHE_LINE; ++i)
EFX_SET_QWORD(entry[i]);
}
#else
/*
* It is possible to do it using AVX2 and AVX512F, but it shows less
* performance.
*/
static inline void
sfc_ef10_ev_qclear_cache_line(void *ptr)
{
const efsys_uint128_t val = _mm_set1_epi64x(UINT64_MAX);
efsys_uint128_t *addr = ptr;
unsigned int i;
RTE_BUILD_BUG_ON(sizeof(val) > RTE_CACHE_LINE_SIZE);
RTE_BUILD_BUG_ON(RTE_CACHE_LINE_SIZE % sizeof(val) != 0);
for (i = 0; i < RTE_CACHE_LINE_SIZE / sizeof(val); ++i)
_mm_store_si128(&addr[i], val);
}
#endif
static inline void
sfc_ef10_ev_qclear(efx_qword_t *hw_ring, unsigned int ptr_mask,
unsigned int old_read_ptr, unsigned int read_ptr)
{
const unsigned int clear_ptr = read_ptr & SFC_EF10_EV_QCLEAR_MASK;
unsigned int old_clear_ptr = old_read_ptr & SFC_EF10_EV_QCLEAR_MASK;
while (old_clear_ptr != clear_ptr) {
sfc_ef10_ev_qclear_cache_line(
&hw_ring[old_clear_ptr & ptr_mask]);
old_clear_ptr += SFC_EF10_EV_PER_CACHE_LINE;
}
/*
* No barriers here.
* Functions which push doorbell should care about correct
* ordering: store instructions which fill in EvQ ring should be
* retired from CPU and DMA sync before doorbell which will allow
* to use these event entries.
*/
}
static inline bool
sfc_ef10_ev_present(const efx_qword_t ev)
{
return ~EFX_QWORD_FIELD(ev, EFX_DWORD_0) |
~EFX_QWORD_FIELD(ev, EFX_DWORD_1);
}
/**
* Alignment requirement for value written to RX WPTR:
* the WPTR must be aligned to an 8 descriptor boundary.
*/
#define SFC_EF10_RX_WPTR_ALIGN 8u
static inline void
sfc_ef10_rx_qpush(volatile void *doorbell, unsigned int added,
unsigned int ptr_mask)
{
efx_dword_t dword;
/* Hardware has alignment restriction for WPTR */
RTE_BUILD_BUG_ON(SFC_RX_REFILL_BULK % SFC_EF10_RX_WPTR_ALIGN != 0);
SFC_ASSERT(RTE_ALIGN(added, SFC_EF10_RX_WPTR_ALIGN) == added);
EFX_POPULATE_DWORD_1(dword, ERF_DZ_RX_DESC_WPTR, added & ptr_mask);
/* DMA sync to device is not required */
/*
* rte_write32() has rte_io_wmb() which guarantees that the STORE
* operations (i.e. Rx and event descriptor updates) that precede
* the rte_io_wmb() call are visible to NIC before the STORE
* operations that follow it (i.e. doorbell write).
*/
rte_write32(dword.ed_u32[0], doorbell);
}
static inline void
sfc_ef10_ev_qprime(volatile void *qprime, unsigned int read_ptr,
unsigned int ptr_mask)
{
efx_dword_t dword;
EFX_POPULATE_DWORD_1(dword, ERF_DZ_EVQ_RPTR, read_ptr & ptr_mask);
rte_write32_relaxed(dword.ed_u32[0], qprime);
rte_wmb();
}
const uint32_t * sfc_ef10_supported_ptypes_get(uint32_t tunnel_encaps);
#ifdef __cplusplus
}
#endif
#endif /* _SFC_EF10_H */