From patchwork Thu Jul 30 10:51:56 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Vladimir Kondratiev X-Patchwork-Id: 6901561 X-Patchwork-Delegate: kvalo@adurom.com Return-Path: X-Original-To: patchwork-linux-wireless@patchwork.kernel.org Delivered-To: patchwork-parsemail@patchwork1.web.kernel.org Received: from mail.kernel.org (mail.kernel.org [198.145.29.136]) by patchwork1.web.kernel.org (Postfix) with ESMTP id 308479F38B for ; Thu, 30 Jul 2015 10:55:34 +0000 (UTC) Received: from mail.kernel.org (localhost [127.0.0.1]) by mail.kernel.org (Postfix) with ESMTP id A4A2C20582 for ; Thu, 30 Jul 2015 10:55:32 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id D400D2056D for ; Thu, 30 Jul 2015 10:55:30 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755542AbbG3KzY (ORCPT ); Thu, 30 Jul 2015 06:55:24 -0400 Received: from sabertooth02.qualcomm.com ([65.197.215.38]:36497 "EHLO sabertooth02.qualcomm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755456AbbG3Kyo (ORCPT ); Thu, 30 Jul 2015 06:54:44 -0400 DKIM-Signature: v=1; a=rsa-sha256; c=simple/simple; d=qca.qualcomm.com; i=@qca.qualcomm.com; q=dns/txt; s=qcdkim; t=1438253684; x=1469789684; h=from:cc:to:subject:date:message-id:in-reply-to: references; bh=wdeIVNBLbRP/VBTWMHETXfQBVHhzNV84SRgWsq2nFV0=; b=uCZoK/IzyUJ6Xg0arJSlh8AfCyvbXNRPxoG+/1eR+SaokATCpY7JFSGl li0y6iMscwOr1yWd+D20SFylLkwu0vGdGQHA0t88qwZOmgd9+KLZR9XGU xuPg2ofVyZXCm72yNiy6U++bGnc8zLuy9LXP4q+J7tjEAwguSKTqygUqh 0=; X-IronPort-AV: E=McAfee;i="5700,7163,7877"; a="94860445" Received: from ironmsg02-r.qualcomm.com ([172.30.46.16]) by sabertooth02.qualcomm.com with ESMTP/TLS/DHE-RSA-AES256-SHA; 30 Jul 2015 03:54:42 -0700 From: Vladimir Kondratiev Cc: Vladimir Kondratiev , linux-wireless@vger.kernel.org, wil6210@qca.qualcomm.com, Vladimir Shulman X-IronPort-AV: E=Sophos;i="5.15,576,1432623600"; d="scan'208";a="523805018" Received: from lx-wigig-72.mea.qualcomm.com ([10.18.134.58]) by ironmsg02-R.qualcomm.com with ESMTP; 30 Jul 2015 03:52:25 -0700 To: Kalle Valo Subject: [PATCH v3 08/19] wil6210: TSO implementation Date: Thu, 30 Jul 2015 13:51:56 +0300 Message-Id: <1438253527-3372-9-git-send-email-qca_vkondrat@qca.qualcomm.com> X-Mailer: git-send-email 2.1.4 In-Reply-To: <1438253527-3372-1-git-send-email-qca_vkondrat@qca.qualcomm.com> References: <1438253527-3372-1-git-send-email-qca_vkondrat@qca.qualcomm.com> Sender: linux-wireless-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-wireless@vger.kernel.org X-Spam-Status: No, score=-8.2 required=5.0 tests=BAYES_00,DKIM_SIGNED, RCVD_IN_DNSWL_HI,RP_MATCHES_RCVD,T_DKIM_INVALID,UNPARSEABLE_RELAY autolearn=unavailable version=3.3.1 X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on mail.kernel.org X-Virus-Scanned: ClamAV using ClamSMTP Driver report supported TSO (v4 & v6) and IP checksum offload in addition to previously supported features. In data path skbs are checked for non-zero gso_size, and when detected sent to additional function for processing TSO SKBs. Since HW does not fully support TSO, additional effort is required from the driver. Driver partitions the data into mss sized descriptors which are then DMAed to the HW. Signed-off-by: Vladimir Shulman Signed-off-by: Vladimir Kondratiev --- drivers/net/wireless/ath/wil6210/netdev.c | 4 +- drivers/net/wireless/ath/wil6210/txrx.c | 371 +++++++++++++++++++++++++++++- drivers/net/wireless/ath/wil6210/txrx.h | 8 + 3 files changed, 371 insertions(+), 12 deletions(-) diff --git a/drivers/net/wireless/ath/wil6210/netdev.c b/drivers/net/wireless/ath/wil6210/netdev.c index 8ef18ac..25c5116 100644 --- a/drivers/net/wireless/ath/wil6210/netdev.c +++ b/drivers/net/wireless/ath/wil6210/netdev.c @@ -173,7 +173,9 @@ void *wil_if_alloc(struct device *dev) wil_set_ethtoolops(ndev); ndev->ieee80211_ptr = wdev; ndev->hw_features = NETIF_F_HW_CSUM | NETIF_F_RXCSUM | - NETIF_F_SG | NETIF_F_GRO; + NETIF_F_SG | NETIF_F_GRO | + NETIF_F_TSO | NETIF_F_TSO6; + ndev->features |= ndev->hw_features; SET_NETDEV_DEV(ndev, wiphy_dev(wdev->wiphy)); wdev->netdev = ndev; diff --git a/drivers/net/wireless/ath/wil6210/txrx.c b/drivers/net/wireless/ath/wil6210/txrx.c index aa20af8..7722df7 100644 --- a/drivers/net/wireless/ath/wil6210/txrx.c +++ b/drivers/net/wireless/ath/wil6210/txrx.c @@ -1058,14 +1058,52 @@ static int wil_tx_desc_map(struct vring_tx_desc *d, dma_addr_t pa, u32 len, static inline void wil_tx_desc_set_nr_frags(struct vring_tx_desc *d, int nr_frags) { - d->mac.d[2] |= ((nr_frags + 1) << - MAC_CFG_DESC_TX_2_NUM_OF_DESCRIPTORS_POS); + d->mac.d[2] |= (nr_frags << MAC_CFG_DESC_TX_2_NUM_OF_DESCRIPTORS_POS); } -static int wil_tx_desc_offload_cksum_set(struct wil6210_priv *wil, - struct vring_tx_desc *d, - struct sk_buff *skb) +/** + * Sets the descriptor @d up for csum and/or TSO offloading. The corresponding + * @skb is used to obtain the protocol and headers length. + * @tso_desc_type is a descriptor type for TSO: 0 - a header, 1 - first data, + * 2 - middle, 3 - last descriptor. + */ + +static void wil_tx_desc_offload_setup_tso(struct vring_tx_desc *d, + struct sk_buff *skb, + int tso_desc_type, bool is_ipv4, + int tcp_hdr_len, int skb_net_hdr_len) { + d->dma.b11 = ETH_HLEN; /* MAC header length */ + d->dma.b11 |= is_ipv4 << DMA_CFG_DESC_TX_OFFLOAD_CFG_L3T_IPV4_POS; + + d->dma.d0 |= (2 << DMA_CFG_DESC_TX_0_L4_TYPE_POS); + /* L4 header len: TCP header length */ + d->dma.d0 |= (tcp_hdr_len & DMA_CFG_DESC_TX_0_L4_LENGTH_MSK); + + /* Setup TSO: bit and desc type */ + d->dma.d0 |= (BIT(DMA_CFG_DESC_TX_0_TCP_SEG_EN_POS)) | + (tso_desc_type << DMA_CFG_DESC_TX_0_SEGMENT_BUF_DETAILS_POS); + d->dma.d0 |= (is_ipv4 << DMA_CFG_DESC_TX_0_IPV4_CHECKSUM_EN_POS); + + d->dma.ip_length = skb_net_hdr_len; + /* Enable TCP/UDP checksum */ + d->dma.d0 |= BIT(DMA_CFG_DESC_TX_0_TCP_UDP_CHECKSUM_EN_POS); + /* Calculate pseudo-header */ + d->dma.d0 |= BIT(DMA_CFG_DESC_TX_0_PSEUDO_HEADER_CALC_EN_POS); +} + +/** + * Sets the descriptor @d up for csum. The corresponding + * @skb is used to obtain the protocol and headers length. + * Returns the protocol: 0 - not TCP, 1 - TCPv4, 2 - TCPv6. + * Note, if d==NULL, the function only returns the protocol result. + * + * It is very similar to previous wil_tx_desc_offload_setup_tso. This + * is "if unrolling" to optimize the critical path. + */ + +static int wil_tx_desc_offload_setup(struct vring_tx_desc *d, + struct sk_buff *skb){ int protocol; if (skb->ip_summed != CHECKSUM_PARTIAL) @@ -1110,6 +1148,305 @@ static int wil_tx_desc_offload_cksum_set(struct wil6210_priv *wil, return 0; } +static inline void wil_tx_last_desc(struct vring_tx_desc *d) +{ + d->dma.d0 |= BIT(DMA_CFG_DESC_TX_0_CMD_EOP_POS) | + BIT(DMA_CFG_DESC_TX_0_CMD_MARK_WB_POS) | + BIT(DMA_CFG_DESC_TX_0_CMD_DMA_IT_POS); +} + +static inline void wil_set_tx_desc_last_tso(volatile struct vring_tx_desc *d) +{ + d->dma.d0 |= wil_tso_type_lst << + DMA_CFG_DESC_TX_0_SEGMENT_BUF_DETAILS_POS; +} + +static int __wil_tx_vring_tso(struct wil6210_priv *wil, struct vring *vring, + struct sk_buff *skb) +{ + struct device *dev = wil_to_dev(wil); + + /* point to descriptors in shared memory */ + volatile struct vring_tx_desc *_desc = NULL, *_hdr_desc, + *_first_desc = NULL; + + /* pointers to shadow descriptors */ + struct vring_tx_desc desc_mem, hdr_desc_mem, first_desc_mem, + *d = &hdr_desc_mem, *hdr_desc = &hdr_desc_mem, + *first_desc = &first_desc_mem; + + /* pointer to shadow descriptors' context */ + struct wil_ctx *hdr_ctx, *first_ctx = NULL; + + int descs_used = 0; /* total number of used descriptors */ + int sg_desc_cnt = 0; /* number of descriptors for current mss*/ + + u32 swhead = vring->swhead; + int used, avail = wil_vring_avail_tx(vring); + int nr_frags = skb_shinfo(skb)->nr_frags; + int min_desc_required = nr_frags + 1; + int mss = skb_shinfo(skb)->gso_size; /* payload size w/o headers */ + int f, len, hdrlen, headlen; + int vring_index = vring - wil->vring_tx; + struct vring_tx_data *txdata = &wil->vring_tx_data[vring_index]; + uint i = swhead; + dma_addr_t pa; + const skb_frag_t *frag = NULL; + int rem_data = mss; + int lenmss; + int hdr_compensation_need = true; + int desc_tso_type = wil_tso_type_first; + bool is_ipv4; + int tcp_hdr_len; + int skb_net_hdr_len; + int gso_type; + + wil_dbg_txrx(wil, "%s() %d bytes to vring %d\n", + __func__, skb->len, vring_index); + + if (unlikely(!txdata->enabled)) + return -EINVAL; + + /* A typical page 4K is 3-4 payloads, we assume each fragment + * is a full payload, that's how min_desc_required has been + * calculated. In real we might need more or less descriptors, + * this is the initial check only. + */ + if (unlikely(avail < min_desc_required)) { + wil_err_ratelimited(wil, + "TSO: Tx ring[%2d] full. No space for %d fragments\n", + vring_index, min_desc_required); + return -ENOMEM; + } + + /* Header Length = MAC header len + IP header len + TCP header len*/ + hdrlen = ETH_HLEN + + (int)skb_network_header_len(skb) + + tcp_hdrlen(skb); + + gso_type = skb_shinfo(skb)->gso_type & (SKB_GSO_TCPV6 | SKB_GSO_TCPV4); + switch (gso_type) { + case SKB_GSO_TCPV4: + /* TCP v4, zero out the IP length and IPv4 checksum fields + * as required by the offloading doc + */ + ip_hdr(skb)->tot_len = 0; + ip_hdr(skb)->check = 0; + is_ipv4 = true; + break; + case SKB_GSO_TCPV6: + /* TCP v6, zero out the payload length */ + ipv6_hdr(skb)->payload_len = 0; + is_ipv4 = false; + break; + default: + /* other than TCPv4 or TCPv6 types are not supported for TSO. + * It is also illegal for both to be set simultaneously + */ + return -EINVAL; + } + + if (skb->ip_summed != CHECKSUM_PARTIAL) + return -EINVAL; + + /* tcp header length and skb network header length are fixed for all + * packet's descriptors - read then once here + */ + tcp_hdr_len = tcp_hdrlen(skb); + skb_net_hdr_len = skb_network_header_len(skb); + + _hdr_desc = &vring->va[i].tx; + + pa = dma_map_single(dev, skb->data, hdrlen, DMA_TO_DEVICE); + if (unlikely(dma_mapping_error(dev, pa))) { + wil_err(wil, "TSO: Skb head DMA map error\n"); + goto err_exit; + } + + wil_tx_desc_map(hdr_desc, pa, hdrlen, vring_index); + wil_tx_desc_offload_setup_tso(hdr_desc, skb, wil_tso_type_hdr, is_ipv4, + tcp_hdr_len, skb_net_hdr_len); + wil_tx_last_desc(hdr_desc); + + vring->ctx[i].mapped_as = wil_mapped_as_single; + hdr_ctx = &vring->ctx[i]; + + descs_used++; + headlen = skb_headlen(skb) - hdrlen; + + for (f = headlen ? -1 : 0; f < nr_frags; f++) { + if (headlen) { + len = headlen; + wil_dbg_txrx(wil, "TSO: process skb head, len %u\n", + len); + } else { + frag = &skb_shinfo(skb)->frags[f]; + len = frag->size; + wil_dbg_txrx(wil, "TSO: frag[%d]: len %u\n", f, len); + } + + while (len) { + wil_dbg_txrx(wil, + "TSO: len %d, rem_data %d, descs_used %d\n", + len, rem_data, descs_used); + + if (descs_used == avail) { + wil_err(wil, "TSO: ring overflow\n"); + goto dma_error; + } + + lenmss = min_t(int, rem_data, len); + i = (swhead + descs_used) % vring->size; + wil_dbg_txrx(wil, "TSO: lenmss %d, i %d\n", lenmss, i); + + if (!headlen) { + pa = skb_frag_dma_map(dev, frag, + frag->size - len, lenmss, + DMA_TO_DEVICE); + vring->ctx[i].mapped_as = wil_mapped_as_page; + } else { + pa = dma_map_single(dev, + skb->data + + skb_headlen(skb) - headlen, + lenmss, + DMA_TO_DEVICE); + vring->ctx[i].mapped_as = wil_mapped_as_single; + headlen -= lenmss; + } + + if (unlikely(dma_mapping_error(dev, pa))) + goto dma_error; + + _desc = &vring->va[i].tx; + + if (!_first_desc) { + _first_desc = _desc; + first_ctx = &vring->ctx[i]; + d = first_desc; + } else { + d = &desc_mem; + } + + wil_tx_desc_map(d, pa, lenmss, vring_index); + wil_tx_desc_offload_setup_tso(d, skb, desc_tso_type, + is_ipv4, tcp_hdr_len, + skb_net_hdr_len); + + /* use tso_type_first only once */ + desc_tso_type = wil_tso_type_mid; + + descs_used++; /* desc used so far */ + sg_desc_cnt++; /* desc used for this segment */ + len -= lenmss; + rem_data -= lenmss; + + wil_dbg_txrx(wil, + "TSO: len %d, rem_data %d, descs_used %d, sg_desc_cnt %d,\n", + len, rem_data, descs_used, sg_desc_cnt); + + /* Close the segment if reached mss size or last frag*/ + if (rem_data == 0 || (f == nr_frags - 1 && len == 0)) { + if (hdr_compensation_need) { + /* first segment include hdr desc for + * release + */ + hdr_ctx->nr_frags = sg_desc_cnt; + wil_tx_desc_set_nr_frags(first_desc, + sg_desc_cnt + + 1); + hdr_compensation_need = false; + } else { + wil_tx_desc_set_nr_frags(first_desc, + sg_desc_cnt); + } + first_ctx->nr_frags = sg_desc_cnt - 1; + + wil_tx_last_desc(d); + + /* first descriptor may also be the last + * for this mss - make sure not to copy + * it twice + */ + if (first_desc != d) + *_first_desc = *first_desc; + + /*last descriptor will be copied at the end + * of this TS processing + */ + if (f < nr_frags - 1 || len > 0) + *_desc = *d; + + rem_data = mss; + _first_desc = NULL; + sg_desc_cnt = 0; + } else if (first_desc != d) /* update mid descriptor */ + *_desc = *d; + } + } + + /* first descriptor may also be the last. + * in this case d pointer is invalid + */ + if (_first_desc == _desc) + d = first_desc; + + /* Last data descriptor */ + wil_set_tx_desc_last_tso(d); + *_desc = *d; + + /* Fill the total number of descriptors in first desc (hdr)*/ + wil_tx_desc_set_nr_frags(hdr_desc, descs_used); + *_hdr_desc = *hdr_desc; + + /* hold reference to skb + * to prevent skb release before accounting + * in case of immediate "tx done" + */ + vring->ctx[i].skb = skb_get(skb); + + /* performance monitoring */ + used = wil_vring_used_tx(vring); + if (wil_val_in_range(vring_idle_trsh, + used, used + descs_used)) { + txdata->idle += get_cycles() - txdata->last_idle; + wil_dbg_txrx(wil, "Ring[%2d] not idle %d -> %d\n", + vring_index, used, used + descs_used); + } + + /* advance swhead */ + wil_dbg_txrx(wil, "TSO: Tx swhead %d -> %d\n", swhead, vring->swhead); + wil_vring_advance_head(vring, descs_used); + + /* make sure all writes to descriptors (shared memory) are done before + * committing them to HW + */ + wmb(); + + iowrite32(vring->swhead, wil->csr + HOSTADDR(vring->hwtail)); + return 0; + +dma_error: + wil_err(wil, "TSO: DMA map page error\n"); + while (descs_used > 0) { + struct wil_ctx *ctx; + + i = (swhead + descs_used) % vring->size; + d = (struct vring_tx_desc *)&vring->va[i].tx; + _desc = &vring->va[i].tx; + *d = *_desc; + _desc->dma.status = TX_DMA_STATUS_DU; + ctx = &vring->ctx[i]; + wil_txdesc_unmap(dev, d, ctx); + if (ctx->skb) + dev_kfree_skb_any(ctx->skb); + memset(ctx, 0, sizeof(*ctx)); + descs_used--; + } + +err_exit: + return -EINVAL; +} + static int __wil_tx_vring(struct wil6210_priv *wil, struct vring *vring, struct sk_buff *skb) { @@ -1128,7 +1465,8 @@ static int __wil_tx_vring(struct wil6210_priv *wil, struct vring *vring, bool mcast = (vring_index == wil->bcast_vring); uint len = skb_headlen(skb); - wil_dbg_txrx(wil, "%s()\n", __func__); + wil_dbg_txrx(wil, "%s() %d bytes to vring %d\n", + __func__, skb->len, vring_index); if (unlikely(!txdata->enabled)) return -EINVAL; @@ -1159,14 +1497,14 @@ static int __wil_tx_vring(struct wil6210_priv *wil, struct vring *vring, d->mac.d[0] |= (1 << MAC_CFG_DESC_TX_0_MCS_INDEX_POS); } /* Process TCP/UDP checksum offloading */ - if (unlikely(wil_tx_desc_offload_cksum_set(wil, d, skb))) { + if (unlikely(wil_tx_desc_offload_setup(d, skb))) { wil_err(wil, "Tx[%2d] Failed to set cksum, drop packet\n", vring_index); goto dma_error; } vring->ctx[i].nr_frags = nr_frags; - wil_tx_desc_set_nr_frags(d, nr_frags); + wil_tx_desc_set_nr_frags(d, nr_frags + 1); /* middle segments */ for (; f < nr_frags; f++) { @@ -1190,7 +1528,7 @@ static int __wil_tx_vring(struct wil6210_priv *wil, struct vring *vring, * if it succeeded for 1-st descriptor, * it will succeed here too */ - wil_tx_desc_offload_cksum_set(wil, d, skb); + wil_tx_desc_offload_setup(d, skb); } /* for the last seg only */ d->dma.d0 |= BIT(DMA_CFG_DESC_TX_0_CMD_EOP_POS); @@ -1221,6 +1559,12 @@ static int __wil_tx_vring(struct wil6210_priv *wil, struct vring *vring, wil_dbg_txrx(wil, "Tx[%2d] swhead %d -> %d\n", vring_index, swhead, vring->swhead); trace_wil6210_tx(vring_index, swhead, skb->len, nr_frags); + + /* make sure all writes to descriptors (shared memory) are done before + * committing them to HW + */ + wmb(); + iowrite32(vring->swhead, wil->csr + HOSTADDR(vring->hwtail)); return 0; @@ -1254,8 +1598,12 @@ static int wil_tx_vring(struct wil6210_priv *wil, struct vring *vring, int rc; spin_lock(&txdata->lock); - rc = __wil_tx_vring(wil, vring, skb); + + rc = (skb_is_gso(skb) ? __wil_tx_vring_tso : __wil_tx_vring) + (wil, vring, skb); + spin_unlock(&txdata->lock); + return rc; } @@ -1382,7 +1730,8 @@ int wil_tx_complete(struct wil6210_priv *wil, int ringid) struct wil_ctx *ctx = &vring->ctx[vring->swtail]; /** * For the fragmented skb, HW will set DU bit only for the - * last fragment. look for it + * last fragment. look for it. + * In TSO the first DU will include hdr desc */ int lf = (vring->swtail + ctx->nr_frags) % vring->size; /* TODO: check we are not past head */ diff --git a/drivers/net/wireless/ath/wil6210/txrx.h b/drivers/net/wireless/ath/wil6210/txrx.h index 0c46384..82a8f9a 100644 --- a/drivers/net/wireless/ath/wil6210/txrx.h +++ b/drivers/net/wireless/ath/wil6210/txrx.h @@ -291,6 +291,14 @@ struct vring_tx_dma { __le16 length; } __packed; +/* TSO type used in dma descriptor d0 bits 11-12 */ +enum { + wil_tso_type_hdr = 0, + wil_tso_type_first = 1, + wil_tso_type_mid = 2, + wil_tso_type_lst = 3, +}; + /* Rx descriptor - MAC part * [dword 0] * bit 0.. 3 : tid:4 The QoS (b3-0) TID Field