DPDK patches and discussions
 help / color / mirror / Atom feed
From: Jerin Jacob <jerinjacobk@gmail.com>
To: pbhagavatula@marvell.com
Cc: jerinj@marvell.com, Shijith Thotton <sthotton@marvell.com>,
	 Nithin Dabilpuram <ndabilpuram@marvell.com>,
	Kiran Kumar K <kirankumark@marvell.com>,
	 Sunil Kumar Kori <skori@marvell.com>,
	Satha Rao <skoteshwar@marvell.com>,
	dev@dpdk.org
Subject: Re: [PATCH v2 3/3] event/cnxk: use WFE in Tx fc wait
Date: Wed, 14 Jun 2023 16:03:00 +0530	[thread overview]
Message-ID: <CALBAE1MiiRbEm77PZ16uXSzpH9s7BA3KTEbrQeg2n+xaNhnqFw@mail.gmail.com> (raw)
In-Reply-To: <20230613092548.1315-3-pbhagavatula@marvell.com>

On Tue, Jun 13, 2023 at 2:56 PM <pbhagavatula@marvell.com> wrote:
>
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
> Use WFE is Tx path when waiting for space in the Tx queue.
> Depending upon the Tx queue contention and size, WFE will
> reduce the cache pressure and power consumption.
> In multi-core scenarios we have observed up to 8W power reduction.
>
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>

Series Applied to dpdk-next-net-eventdev/for-main. Thanks

> ---
>  drivers/event/cnxk/cn10k_tx_worker.h |  18 ++++
>  drivers/net/cnxk/cn10k_tx.h          | 152 +++++++++++++++++++++++----
>  2 files changed, 147 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/event/cnxk/cn10k_tx_worker.h b/drivers/event/cnxk/cn10k_tx_worker.h
> index b6c9bb1d26..dea6cdcde2 100644
> --- a/drivers/event/cnxk/cn10k_tx_worker.h
> +++ b/drivers/event/cnxk/cn10k_tx_worker.h
> @@ -24,9 +24,27 @@ cn10k_sso_hws_xtract_meta(struct rte_mbuf *m, const uint64_t *txq_data)
>  static __rte_always_inline void
>  cn10k_sso_txq_fc_wait(const struct cn10k_eth_txq *txq)
>  {
> +#ifdef RTE_ARCH_ARM64
> +       uint64_t space;
> +
> +       asm volatile(PLT_CPU_FEATURE_PREAMBLE
> +                    "          ldxr %[space], [%[addr]]                \n"
> +                    "          cmp %[adj], %[space]                    \n"
> +                    "          b.hi .Ldne%=                            \n"
> +                    "          sevl                                    \n"
> +                    ".Lrty%=:  wfe                                     \n"
> +                    "          ldxr %[space], [%[addr]]                \n"
> +                    "          cmp %[adj], %[space]                    \n"
> +                    "          b.ls .Lrty%=                            \n"
> +                    ".Ldne%=:                                          \n"
> +                    : [space] "=&r"(space)
> +                    : [adj] "r"(txq->nb_sqb_bufs_adj), [addr] "r"(txq->fc_mem)
> +                    : "memory");
> +#else
>         while ((uint64_t)txq->nb_sqb_bufs_adj <=
>                __atomic_load_n(txq->fc_mem, __ATOMIC_RELAXED))
>                 ;
> +#endif
>  }
>
>  static __rte_always_inline int32_t
> diff --git a/drivers/net/cnxk/cn10k_tx.h b/drivers/net/cnxk/cn10k_tx.h
> index a365cbe0ee..d0e8350ce2 100644
> --- a/drivers/net/cnxk/cn10k_tx.h
> +++ b/drivers/net/cnxk/cn10k_tx.h
> @@ -102,27 +102,72 @@ cn10k_nix_tx_mbuf_validate(struct rte_mbuf *m, const uint32_t flags)
>  }
>
>  static __plt_always_inline void
> -cn10k_nix_vwqe_wait_fc(struct cn10k_eth_txq *txq, int64_t req)
> +cn10k_nix_vwqe_wait_fc(struct cn10k_eth_txq *txq, uint16_t req)
>  {
>         int64_t cached, refill;
> +       int64_t pkts;
>
>  retry:
> +#ifdef RTE_ARCH_ARM64
> +
> +       asm volatile(PLT_CPU_FEATURE_PREAMBLE
> +                    "          ldxr %[pkts], [%[addr]]                 \n"
> +                    "          tbz %[pkts], 63, .Ldne%=                \n"
> +                    "          sevl                                    \n"
> +                    ".Lrty%=:  wfe                                     \n"
> +                    "          ldxr %[pkts], [%[addr]]                 \n"
> +                    "          tbnz %[pkts], 63, .Lrty%=               \n"
> +                    ".Ldne%=:                                          \n"
> +                    : [pkts] "=&r"(pkts)
> +                    : [addr] "r"(&txq->fc_cache_pkts)
> +                    : "memory");
> +#else
> +       RTE_SET_USED(pkts);
>         while (__atomic_load_n(&txq->fc_cache_pkts, __ATOMIC_RELAXED) < 0)
>                 ;
> +#endif
>         cached = __atomic_fetch_sub(&txq->fc_cache_pkts, req, __ATOMIC_ACQUIRE) - req;
>         /* Check if there is enough space, else update and retry. */
> -       if (cached < 0) {
> -               /* Check if we have space else retry. */
> -               do {
> -                       refill = txq->nb_sqb_bufs_adj -
> -                                __atomic_load_n(txq->fc_mem, __ATOMIC_RELAXED);
> -                       refill = (refill << txq->sqes_per_sqb_log2) - refill;
> -               } while (refill <= 0);
> -               __atomic_compare_exchange(&txq->fc_cache_pkts, &cached, &refill,
> -                                         0, __ATOMIC_RELEASE,
> -                                         __ATOMIC_RELAXED);
> +       if (cached >= 0)
> +               return;
> +
> +       /* Check if we have space else retry. */
> +#ifdef RTE_ARCH_ARM64
> +       int64_t val;
> +
> +       asm volatile(PLT_CPU_FEATURE_PREAMBLE
> +                    "          ldxr %[val], [%[addr]]                  \n"
> +                    "          sub %[val], %[adj], %[val]              \n"
> +                    "          lsl %[refill], %[val], %[shft]          \n"
> +                    "          sub %[refill], %[refill], %[val]        \n"
> +                    "          sub %[refill], %[refill], %[sub]        \n"
> +                    "          cmp %[refill], #0x0                     \n"
> +                    "          b.ge .Ldne%=                            \n"
> +                    "          sevl                                    \n"
> +                    ".Lrty%=:  wfe                                     \n"
> +                    "          ldxr %[val], [%[addr]]                  \n"
> +                    "          sub %[val], %[adj], %[val]              \n"
> +                    "          lsl %[refill], %[val], %[shft]          \n"
> +                    "          sub %[refill], %[refill], %[val]        \n"
> +                    "          sub %[refill], %[refill], %[sub]        \n"
> +                    "          cmp %[refill], #0x0                     \n"
> +                    "          b.lt .Lrty%=                            \n"
> +                    ".Ldne%=:                                          \n"
> +                    : [refill] "=&r"(refill), [val] "=&r" (val)
> +                    : [addr] "r"(txq->fc_mem), [adj] "r"(txq->nb_sqb_bufs_adj),
> +                      [shft] "r"(txq->sqes_per_sqb_log2), [sub] "r"(req)
> +                    : "memory");
> +#else
> +       do {
> +               refill = (txq->nb_sqb_bufs_adj - __atomic_load_n(txq->fc_mem, __ATOMIC_RELAXED));
> +               refill = (refill << txq->sqes_per_sqb_log2) - refill;
> +               refill -= req;
> +       } while (refill < 0);
> +#endif
> +       if (!__atomic_compare_exchange(&txq->fc_cache_pkts, &cached, &refill,
> +                                 0, __ATOMIC_RELEASE,
> +                                 __ATOMIC_RELAXED))
>                 goto retry;
> -       }
>  }
>
>  /* Function to determine no of tx subdesc required in case ext
> @@ -283,10 +328,27 @@ static __rte_always_inline void
>  cn10k_nix_sec_fc_wait_one(struct cn10k_eth_txq *txq)
>  {
>         uint64_t nb_desc = txq->cpt_desc;
> -       uint64_t *fc = txq->cpt_fc;
> -
> -       while (nb_desc <= __atomic_load_n(fc, __ATOMIC_RELAXED))
> +       uint64_t fc;
> +
> +#ifdef RTE_ARCH_ARM64
> +       asm volatile(PLT_CPU_FEATURE_PREAMBLE
> +                    "          ldxr %[space], [%[addr]]                \n"
> +                    "          cmp %[nb_desc], %[space]                \n"
> +                    "          b.hi .Ldne%=                            \n"
> +                    "          sevl                                    \n"
> +                    ".Lrty%=:  wfe                                     \n"
> +                    "          ldxr %[space], [%[addr]]                \n"
> +                    "          cmp %[nb_desc], %[space]                \n"
> +                    "          b.ls .Lrty%=                            \n"
> +                    ".Ldne%=:                                          \n"
> +                    : [space] "=&r"(fc)
> +                    : [nb_desc] "r"(nb_desc), [addr] "r"(txq->cpt_fc)
> +                    : "memory");
> +#else
> +       RTE_SET_USED(fc);
> +       while (nb_desc <= __atomic_load_n(txq->cpt_fc, __ATOMIC_RELAXED))
>                 ;
> +#endif
>  }
>
>  static __rte_always_inline void
> @@ -294,7 +356,7 @@ cn10k_nix_sec_fc_wait(struct cn10k_eth_txq *txq, uint16_t nb_pkts)
>  {
>         int32_t nb_desc, val, newval;
>         int32_t *fc_sw;
> -       volatile uint64_t *fc;
> +       uint64_t *fc;
>
>         /* Check if there is any CPT instruction to submit */
>         if (!nb_pkts)
> @@ -302,21 +364,59 @@ cn10k_nix_sec_fc_wait(struct cn10k_eth_txq *txq, uint16_t nb_pkts)
>
>  again:
>         fc_sw = txq->cpt_fc_sw;
> -       val = __atomic_fetch_sub(fc_sw, nb_pkts, __ATOMIC_RELAXED) - nb_pkts;
> +#ifdef RTE_ARCH_ARM64
> +       asm volatile(PLT_CPU_FEATURE_PREAMBLE
> +                    "          ldxr %w[pkts], [%[addr]]                \n"
> +                    "          tbz %w[pkts], 31, .Ldne%=               \n"
> +                    "          sevl                                    \n"
> +                    ".Lrty%=:  wfe                                     \n"
> +                    "          ldxr %w[pkts], [%[addr]]                \n"
> +                    "          tbnz %w[pkts], 31, .Lrty%=              \n"
> +                    ".Ldne%=:                                          \n"
> +                    : [pkts] "=&r"(val)
> +                    : [addr] "r"(fc_sw)
> +                    : "memory");
> +#else
> +       /* Wait for primary core to refill FC. */
> +       while (__atomic_load_n(fc_sw, __ATOMIC_RELAXED) < 0)
> +               ;
> +#endif
> +
> +       val = __atomic_fetch_sub(fc_sw, nb_pkts, __ATOMIC_ACQUIRE) - nb_pkts;
>         if (likely(val >= 0))
>                 return;
>
>         nb_desc = txq->cpt_desc;
>         fc = txq->cpt_fc;
> +#ifdef RTE_ARCH_ARM64
> +       asm volatile(PLT_CPU_FEATURE_PREAMBLE
> +                    "          ldxr %[refill], [%[addr]]               \n"
> +                    "          sub %[refill], %[desc], %[refill]       \n"
> +                    "          sub %[refill], %[refill], %[pkts]       \n"
> +                    "          cmp %[refill], #0x0                     \n"
> +                    "          b.ge .Ldne%=                            \n"
> +                    "          sevl                                    \n"
> +                    ".Lrty%=:  wfe                                     \n"
> +                    "          ldxr %[refill], [%[addr]]               \n"
> +                    "          sub %[refill], %[desc], %[refill]       \n"
> +                    "          sub %[refill], %[refill], %[pkts]       \n"
> +                    "          cmp %[refill], #0x0                     \n"
> +                    "          b.lt .Lrty%=                            \n"
> +                    ".Ldne%=:                                          \n"
> +                    : [refill] "=&r"(newval)
> +                    : [addr] "r"(fc), [desc] "r"(nb_desc), [pkts] "r"(nb_pkts)
> +                    : "memory");
> +#else
>         while (true) {
>                 newval = nb_desc - __atomic_load_n(fc, __ATOMIC_RELAXED);
>                 newval -= nb_pkts;
>                 if (newval >= 0)
>                         break;
>         }
> +#endif
>
> -       if (!__atomic_compare_exchange_n(fc_sw, &val, newval, false,
> -                                        __ATOMIC_RELAXED, __ATOMIC_RELAXED))
> +       if (!__atomic_compare_exchange_n(fc_sw, &val, newval, false, __ATOMIC_RELEASE,
> +                                        __ATOMIC_RELAXED))
>                 goto again;
>  }
>
> @@ -3033,10 +3133,16 @@ cn10k_nix_xmit_pkts_vector(void *tx_queue, uint64_t *ws,
>                 wd.data[1] |= ((uint64_t)(lnum - 17)) << 12;
>                 wd.data[1] |= (uint64_t)(lmt_id + 16);
>
> -               if (flags & NIX_TX_VWQE_F)
> -                       cn10k_nix_vwqe_wait_fc(txq,
> -                               burst - (cn10k_nix_pkts_per_vec_brst(flags) >>
> -                                        1));
> +               if (flags & NIX_TX_VWQE_F) {
> +                       if (flags & NIX_TX_MULTI_SEG_F) {
> +                               if (burst - (cn10k_nix_pkts_per_vec_brst(flags) >> 1) > 0)
> +                                       cn10k_nix_vwqe_wait_fc(txq,
> +                                               burst - (cn10k_nix_pkts_per_vec_brst(flags) >> 1));
> +                       } else {
> +                               cn10k_nix_vwqe_wait_fc(txq,
> +                                               burst - (cn10k_nix_pkts_per_vec_brst(flags) >> 1));
> +                       }
> +               }
>                 /* STEOR1 */
>                 roc_lmt_submit_steorl(wd.data[1], pa);
>         } else if (lnum) {
> --
> 2.25.1
>

  reply	other threads:[~2023-06-14 10:33 UTC|newest]

Thread overview: 13+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-05-16 14:37 [PATCH 1/3] event/cnxk: align TX queue buffer adjustment pbhagavatula
2023-05-16 14:37 ` [PATCH 2/3] event/cnxk: use local labels in asm intrinsic pbhagavatula
2023-05-16 14:37 ` [PATCH 3/3] event/cnxk: use WFE in Tx fc wait pbhagavatula
2023-06-12 15:52 ` [PATCH 1/3] event/cnxk: align TX queue buffer adjustment Jerin Jacob
2023-06-13  9:25 ` [PATCH v2 " pbhagavatula
2023-06-13  9:25   ` [PATCH v2 2/3] event/cnxk: use local labels in asm intrinsic pbhagavatula
2023-06-13  9:25   ` [PATCH v2 3/3] event/cnxk: use WFE in Tx fc wait pbhagavatula
2023-06-14 10:33     ` Jerin Jacob [this message]
2023-06-14 18:27       ` Patrick Robb
2023-06-14 20:24         ` [EXT] " Pavan Nikhilesh Bhagavatula
2023-06-15  5:49         ` Jerin Jacob Kollanukkaran
2023-06-15 15:28     ` Stephen Hemminger
2023-06-16  6:46       ` [EXT] " Pavan Nikhilesh Bhagavatula

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CALBAE1MiiRbEm77PZ16uXSzpH9s7BA3KTEbrQeg2n+xaNhnqFw@mail.gmail.com \
    --to=jerinjacobk@gmail.com \
    --cc=dev@dpdk.org \
    --cc=jerinj@marvell.com \
    --cc=kirankumark@marvell.com \
    --cc=ndabilpuram@marvell.com \
    --cc=pbhagavatula@marvell.com \
    --cc=skori@marvell.com \
    --cc=skoteshwar@marvell.com \
    --cc=sthotton@marvell.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).