numam-dpdk/drivers/event/octeontx2/otx2_worker.h
Pavan Nikhilesh ea8403d084 event/octeontx2: remove hot spots by prefetching data
Recent profiling revealed few hotspots in octeontx2 event device
driver add prefetch hints to reduce stalls.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
2019-10-31 11:48:44 +01:00

305 lines
8.1 KiB
C

/* SPDX-License-Identifier: BSD-3-Clause
* Copyright(C) 2019 Marvell International Ltd.
*/
#ifndef __OTX2_WORKER_H__
#define __OTX2_WORKER_H__
#include <rte_common.h>
#include <rte_branch_prediction.h>
#include <otx2_common.h>
#include "otx2_evdev.h"
/* SSO Operations */
static __rte_always_inline uint16_t
otx2_ssogws_get_work(struct otx2_ssogws *ws, struct rte_event *ev,
const uint32_t flags, const void * const lookup_mem)
{
union otx2_sso_event event;
uint64_t tstamp_ptr;
uint64_t get_work1;
uint64_t mbuf;
otx2_write64(BIT_ULL(16) | /* wait for work. */
1, /* Use Mask set 0. */
ws->getwrk_op);
if (flags & NIX_RX_OFFLOAD_PTYPE_F)
rte_prefetch_non_temporal(lookup_mem);
#ifdef RTE_ARCH_ARM64
asm volatile(
" ldr %[tag], [%[tag_loc]] \n"
" ldr %[wqp], [%[wqp_loc]] \n"
" tbz %[tag], 63, done%= \n"
" sevl \n"
"rty%=: wfe \n"
" ldr %[tag], [%[tag_loc]] \n"
" ldr %[wqp], [%[wqp_loc]] \n"
" tbnz %[tag], 63, rty%= \n"
"done%=: dmb ld \n"
" prfm pldl1keep, [%[wqp], #8] \n"
" sub %[mbuf], %[wqp], #0x80 \n"
" prfm pldl1keep, [%[mbuf]] \n"
: [tag] "=&r" (event.get_work0),
[wqp] "=&r" (get_work1),
[mbuf] "=&r" (mbuf)
: [tag_loc] "r" (ws->tag_op),
[wqp_loc] "r" (ws->wqp_op)
);
#else
event.get_work0 = otx2_read64(ws->tag_op);
while ((BIT_ULL(63)) & event.get_work0)
event.get_work0 = otx2_read64(ws->tag_op);
get_work1 = otx2_read64(ws->wqp_op);
rte_prefetch0((const void *)get_work1);
mbuf = (uint64_t)((char *)get_work1 - sizeof(struct rte_mbuf));
rte_prefetch0((const void *)mbuf);
#endif
event.get_work0 = (event.get_work0 & (0x3ull << 32)) << 6 |
(event.get_work0 & (0x3FFull << 36)) << 4 |
(event.get_work0 & 0xffffffff);
ws->cur_tt = event.sched_type;
ws->cur_grp = event.queue_id;
if (event.sched_type != SSO_TT_EMPTY &&
event.event_type == RTE_EVENT_TYPE_ETHDEV) {
otx2_wqe_to_mbuf(get_work1, mbuf, event.sub_event_type,
(uint32_t) event.get_work0, flags, lookup_mem);
/* Extracting tstamp, if PTP enabled*/
tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)get_work1)
+ OTX2_SSO_WQE_SG_PTR);
otx2_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, ws->tstamp,
flags, (uint64_t *)tstamp_ptr);
get_work1 = mbuf;
}
ev->event = event.get_work0;
ev->u64 = get_work1;
return !!get_work1;
}
/* Used in cleaning up workslot. */
static __rte_always_inline uint16_t
otx2_ssogws_get_work_empty(struct otx2_ssogws *ws, struct rte_event *ev,
const uint32_t flags)
{
union otx2_sso_event event;
uint64_t tstamp_ptr;
uint64_t get_work1;
uint64_t mbuf;
#ifdef RTE_ARCH_ARM64
asm volatile(
" ldr %[tag], [%[tag_loc]] \n"
" ldr %[wqp], [%[wqp_loc]] \n"
" tbz %[tag], 63, done%= \n"
" sevl \n"
"rty%=: wfe \n"
" ldr %[tag], [%[tag_loc]] \n"
" ldr %[wqp], [%[wqp_loc]] \n"
" tbnz %[tag], 63, rty%= \n"
"done%=: dmb ld \n"
" prfm pldl1keep, [%[wqp], #8] \n"
" sub %[mbuf], %[wqp], #0x80 \n"
" prfm pldl1keep, [%[mbuf]] \n"
: [tag] "=&r" (event.get_work0),
[wqp] "=&r" (get_work1),
[mbuf] "=&r" (mbuf)
: [tag_loc] "r" (ws->tag_op),
[wqp_loc] "r" (ws->wqp_op)
);
#else
event.get_work0 = otx2_read64(ws->tag_op);
while ((BIT_ULL(63)) & event.get_work0)
event.get_work0 = otx2_read64(ws->tag_op);
get_work1 = otx2_read64(ws->wqp_op);
rte_prefetch_non_temporal((const void *)get_work1);
mbuf = (uint64_t)((char *)get_work1 - sizeof(struct rte_mbuf));
rte_prefetch_non_temporal((const void *)mbuf);
#endif
event.get_work0 = (event.get_work0 & (0x3ull << 32)) << 6 |
(event.get_work0 & (0x3FFull << 36)) << 4 |
(event.get_work0 & 0xffffffff);
ws->cur_tt = event.sched_type;
ws->cur_grp = event.queue_id;
if (event.sched_type != SSO_TT_EMPTY &&
event.event_type == RTE_EVENT_TYPE_ETHDEV) {
otx2_wqe_to_mbuf(get_work1, mbuf, event.sub_event_type,
(uint32_t) event.get_work0, flags, NULL);
/* Extracting tstamp, if PTP enabled*/
tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)get_work1)
+ OTX2_SSO_WQE_SG_PTR);
otx2_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, ws->tstamp,
flags, (uint64_t *)tstamp_ptr);
get_work1 = mbuf;
}
ev->event = event.get_work0;
ev->u64 = get_work1;
return !!get_work1;
}
static __rte_always_inline void
otx2_ssogws_add_work(struct otx2_ssogws *ws, const uint64_t event_ptr,
const uint32_t tag, const uint8_t new_tt,
const uint16_t grp)
{
uint64_t add_work0;
add_work0 = tag | ((uint64_t)(new_tt) << 32);
otx2_store_pair(add_work0, event_ptr, ws->grps_base[grp]);
}
static __rte_always_inline void
otx2_ssogws_swtag_desched(struct otx2_ssogws *ws, uint32_t tag, uint8_t new_tt,
uint16_t grp)
{
uint64_t val;
val = tag | ((uint64_t)(new_tt & 0x3) << 32) | ((uint64_t)grp << 34);
otx2_write64(val, ws->swtag_desched_op);
}
static __rte_always_inline void
otx2_ssogws_swtag_norm(struct otx2_ssogws *ws, uint32_t tag, uint8_t new_tt)
{
uint64_t val;
val = tag | ((uint64_t)(new_tt & 0x3) << 32);
otx2_write64(val, ws->swtag_norm_op);
}
static __rte_always_inline void
otx2_ssogws_swtag_untag(struct otx2_ssogws *ws)
{
otx2_write64(0, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
SSOW_LF_GWS_OP_SWTAG_UNTAG);
ws->cur_tt = SSO_SYNC_UNTAGGED;
}
static __rte_always_inline void
otx2_ssogws_swtag_flush(struct otx2_ssogws *ws)
{
otx2_write64(0, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
SSOW_LF_GWS_OP_SWTAG_FLUSH);
ws->cur_tt = SSO_SYNC_EMPTY;
}
static __rte_always_inline void
otx2_ssogws_desched(struct otx2_ssogws *ws)
{
otx2_write64(0, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
SSOW_LF_GWS_OP_DESCHED);
}
static __rte_always_inline void
otx2_ssogws_swtag_wait(struct otx2_ssogws *ws)
{
#ifdef RTE_ARCH_ARM64
uint64_t swtp;
asm volatile (
" ldr %[swtb], [%[swtp_loc]] \n"
" cbz %[swtb], done%= \n"
" sevl \n"
"rty%=: wfe \n"
" ldr %[swtb], [%[swtp_loc]] \n"
" cbnz %[swtb], rty%= \n"
"done%=: \n"
: [swtb] "=&r" (swtp)
: [swtp_loc] "r" (ws->swtp_op)
);
#else
/* Wait for the SWTAG/SWTAG_FULL operation */
while (otx2_read64(ws->swtp_op))
;
#endif
}
static __rte_always_inline void
otx2_ssogws_head_wait(struct otx2_ssogws *ws)
{
#ifdef RTE_ARCH_ARM64
uint64_t tag;
asm volatile (
" ldr %[tag], [%[tag_op]] \n"
" tbnz %[tag], 35, done%= \n"
" sevl \n"
"rty%=: wfe \n"
" ldr %[tag], [%[tag_op]] \n"
" tbz %[tag], 35, rty%= \n"
"done%=: \n"
: [tag] "=&r" (tag)
: [tag_op] "r" (ws->tag_op)
);
#else
/* Wait for the HEAD to be set */
while (!(otx2_read64(ws->tag_op) & BIT_ULL(35)))
;
#endif
}
static __rte_always_inline void
otx2_ssogws_order(struct otx2_ssogws *ws, const uint8_t wait_flag)
{
if (wait_flag)
otx2_ssogws_head_wait(ws);
rte_cio_wmb();
}
static __rte_always_inline const struct otx2_eth_txq *
otx2_ssogws_xtract_meta(struct rte_mbuf *m)
{
return rte_eth_devices[m->port].data->tx_queues[
rte_event_eth_tx_adapter_txq_get(m)];
}
static __rte_always_inline void
otx2_ssogws_prepare_pkt(const struct otx2_eth_txq *txq, struct rte_mbuf *m,
uint64_t *cmd, const uint32_t flags)
{
otx2_lmt_mov(cmd, txq->cmd, otx2_nix_tx_ext_subs(flags));
otx2_nix_xmit_prepare(m, cmd, flags);
}
static __rte_always_inline uint16_t
otx2_ssogws_event_tx(struct otx2_ssogws *ws, struct rte_event ev[],
uint64_t *cmd, const uint32_t flags)
{
struct rte_mbuf *m = ev[0].mbuf;
const struct otx2_eth_txq *txq = otx2_ssogws_xtract_meta(m);
rte_prefetch_non_temporal(txq);
/* Perform header writes before barrier for TSO */
otx2_nix_xmit_prepare_tso(m, flags);
otx2_ssogws_order(ws, !ev->sched_type);
otx2_ssogws_prepare_pkt(txq, m, cmd, flags);
if (flags & NIX_TX_MULTI_SEG_F) {
const uint16_t segdw = otx2_nix_prepare_mseg(m, cmd, flags);
otx2_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
m->ol_flags, segdw, flags);
otx2_nix_xmit_mseg_one(cmd, txq->lmt_addr, txq->io_addr, segdw);
} else {
/* Passing no of segdw as 4: HDR + EXT + SG + SMEM */
otx2_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
m->ol_flags, 4, flags);
otx2_nix_xmit_one(cmd, txq->lmt_addr, txq->io_addr, flags);
}
return 1;
}
#endif