package/kernel/mac80211/patches/358-mac80211-add-NEED_ALIGNED4_SKBS-hw-flag.patch

   1 From: Janusz Dziedzic <janusz.dziedzic@tieto.com>
   2 Date: Fri, 19 Feb 2016 11:01:50 +0100
   3 Subject: [PATCH] mac80211: add NEED_ALIGNED4_SKBS hw flag
   4
   5 HW/driver should set NEED_ALIGNED4_SKBS flag in case
   6 require aligned skbs to four-byte boundaries.
   7 This affect only TX direction.
   8
   9 Padding is added after ieee80211_hdr, before IV/LLC.
  10
  11 Before we have to do memmove(hdrlen) twice in the
  12 dirver. Once before we pass this to HW and next
  13 in tx completion (to be sure monitor will report
  14 this tx frame correctly).
  15
  16 With this patch we can skip this memmove() and save CPU.
  17
  18 Currently this was tested with ath9k, both hw/sw crypt for
  19 wep/tkip/ccmp.
  20
  21 Signed-off-by: Janusz Dziedzic <janusz.dziedzic@tieto.com>
  22 ---
  23
  24 --- a/include/net/mac80211.h
  25 +++ b/include/net/mac80211.h
  26 @@ -2059,6 +2059,9 @@ struct ieee80211_txq {
  27   * @IEEE80211_HW_SUPPORTS_TDLS_BUFFER_STA: Hardware supports buffer STA on
  28   *     TDLS links.
  29   *
  30 + * @IEEE80211_HW_NEEDS_ALIGNED4_SKBS: Driver need aligned skbs to four-byte.
  31 + *     Padding will be added after ieee80211_hdr, before IV/LLC.
  32 + *
  33   * @NUM_IEEE80211_HW_FLAGS: number of hardware flags, used for sizing arrays
  34   */
  35  enum ieee80211_hw_flags {
  36 @@ -2102,6 +2105,7 @@ enum ieee80211_hw_flags {
  37         IEEE80211_HW_REPORTS_LOW_ACK,
  38         IEEE80211_HW_SUPPORTS_TX_FRAG,
  39         IEEE80211_HW_SUPPORTS_TDLS_BUFFER_STA,
  40 +       IEEE80211_HW_NEEDS_ALIGNED4_SKBS,
  41
  42         /* keep last, obviously */
  43         NUM_IEEE80211_HW_FLAGS
  44 --- a/net/mac80211/debugfs.c
  45 +++ b/net/mac80211/debugfs.c
  46 @@ -212,6 +212,7 @@ static const char *hw_flag_names[] = {
  47         FLAG(REPORTS_LOW_ACK),
  48         FLAG(SUPPORTS_TX_FRAG),
  49         FLAG(SUPPORTS_TDLS_BUFFER_STA),
  50 +       FLAG(NEEDS_ALIGNED4_SKBS),
  51  #undef FLAG
  52  };
  53
  54 --- a/net/mac80211/ieee80211_i.h
  55 +++ b/net/mac80211/ieee80211_i.h
  56 @@ -1548,6 +1548,29 @@ ieee80211_vif_get_num_mcast_if(struct ie
  57         return -1;
  58  }
  59
  60 +static inline unsigned int
  61 +ieee80211_hdr_padsize(struct ieee80211_hw *hw, unsigned int hdrlen)
  62 +{
  63 +       /*
  64 +        * While hdrlen is already aligned to two-byte boundaries,
  65 +        * simple check with & 2 will return correct padsize.
  66 +        */
  67 +       if (ieee80211_hw_check(hw, NEEDS_ALIGNED4_SKBS))
  68 +               return hdrlen & 2;
  69 +       return 0;
  70 +}
  71 +
  72 +static inline unsigned int
  73 +ieee80211_padded_hdrlen(struct ieee80211_hw *hw, __le16 fc)
  74 +{
  75 +       unsigned int hdrlen;
  76 +
  77 +       hdrlen = ieee80211_hdrlen(fc);
  78 +       hdrlen += ieee80211_hdr_padsize(hw, hdrlen);
  79 +
  80 +       return hdrlen;
  81 +}
  82 +
  83  u64 ieee80211_calculate_rx_timestamp(struct ieee80211_local *local,
  84                                      struct ieee80211_rx_status *status,
  85                                      unsigned int mpdu_len,
  86 --- a/net/mac80211/sta_info.h
  87 +++ b/net/mac80211/sta_info.h
  88 @@ -300,7 +300,7 @@ struct ieee80211_fast_tx {
  89         u8 hdr_len;
  90         u8 sa_offs, da_offs, pn_offs;
  91         u8 band;
  92 -       u8 hdr[30 + 2 + IEEE80211_FAST_XMIT_MAX_IV +
  93 +       u8 hdr[30 + 2 + 2 + IEEE80211_FAST_XMIT_MAX_IV +
  94                sizeof(rfc1042_header)] __aligned(2);
  95
  96         struct rcu_head rcu_head;
  97 --- a/net/mac80211/status.c
  98 +++ b/net/mac80211/status.c
  99 @@ -642,9 +642,22 @@ void ieee80211_tx_monitor(struct ieee802
 100         struct sk_buff *skb2;
 101         struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
 102         struct ieee80211_sub_if_data *sdata;
 103 +       struct ieee80211_hdr *hdr = (void *)skb->data;
 104         struct net_device *prev_dev = NULL;
 105 +       unsigned int hdrlen, padsize;
 106         int rtap_len;
 107
 108 +       /* Remove padding if was added */
 109 +       if (ieee80211_hw_check(&local->hw, NEEDS_ALIGNED4_SKBS)) {
 110 +               hdrlen = ieee80211_hdrlen(hdr->frame_control);
 111 +               padsize = ieee80211_hdr_padsize(&local->hw, hdrlen);
 112 +
 113 +               if (padsize && skb->len > hdrlen + padsize) {
 114 +                       memmove(skb->data + padsize, skb->data, hdrlen);
 115 +                       skb_pull(skb, padsize);
 116 +               }
 117 +       }
 118 +
 119         /* send frame to monitor interfaces now */
 120         rtap_len = ieee80211_tx_radiotap_len(info);
 121         if (WARN_ON_ONCE(skb_headroom(skb) < rtap_len)) {
 122 --- a/net/mac80211/tkip.c
 123 +++ b/net/mac80211/tkip.c
 124 @@ -201,10 +201,12 @@ void ieee80211_get_tkip_p2k(struct ieee8
 125  {
 126         struct ieee80211_key *key = (struct ieee80211_key *)
 127                         container_of(keyconf, struct ieee80211_key, conf);
 128 +       struct ieee80211_hw *hw = &key->local->hw;
 129         const u8 *tk = &key->conf.key[NL80211_TKIP_DATA_OFFSET_ENCR_KEY];
 130         struct tkip_ctx *ctx = &key->u.tkip.tx;
 131         struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)skb->data;
 132 -       const u8 *data = (u8 *)hdr + ieee80211_hdrlen(hdr->frame_control);
 133 +       const u8 *data = (u8 *)hdr + ieee80211_padded_hdrlen(hw,
 134 +                                                       hdr->frame_control);
 135         u32 iv32 = get_unaligned_le32(&data[4]);
 136         u16 iv16 = data[2] | (data[0] << 8);
 137
 138 --- a/net/mac80211/tx.c
 139 +++ b/net/mac80211/tx.c
 140 @@ -1172,8 +1172,7 @@ ieee80211_tx_prepare(struct ieee80211_su
 141         info->flags &= ~IEEE80211_TX_INTFL_NEED_TXPROCESSING;
 142
 143         hdr = (struct ieee80211_hdr *) skb->data;
 144 -
 145 -       tx->hdrlen = ieee80211_hdrlen(hdr->frame_control);
 146 +       tx->hdrlen = ieee80211_padded_hdrlen(&local->hw, hdr->frame_control);
 147
 148         if (likely(sta)) {
 149                 if (!IS_ERR(sta))
 150 @@ -2184,7 +2183,7 @@ netdev_tx_t ieee80211_monitor_start_xmit
 151                 goto fail;
 152
 153         hdr = (struct ieee80211_hdr *)(skb->data + len_rthdr);
 154 -       hdrlen = ieee80211_hdrlen(hdr->frame_control);
 155 +       hdrlen = ieee80211_padded_hdrlen(&local->hw, hdr->frame_control);
 156
 157         if (skb->len < len_rthdr + hdrlen)
 158                 goto fail;
 159 @@ -2402,7 +2401,7 @@ static struct sk_buff *ieee80211_build_h
 160         struct ieee80211_chanctx_conf *chanctx_conf;
 161         struct ieee80211_sub_if_data *ap_sdata;
 162         enum nl80211_band band;
 163 -       int ret;
 164 +       int padsize, ret;
 165
 166         if (IS_ERR(sta))
 167                 sta = NULL;
 168 @@ -2622,6 +2621,9 @@ static struct sk_buff *ieee80211_build_h
 169                 hdrlen += 2;
 170         }
 171
 172 +       /* Check aligned4 skb required */
 173 +       padsize = ieee80211_hdr_padsize(&sdata->local->hw, hdrlen);
 174 +
 175         /*
 176          * Drop unicast frames to unauthorised stations unless they are
 177          * EAPOL frames from the local station.
 178 @@ -2702,6 +2704,7 @@ static struct sk_buff *ieee80211_build_h
 179
 180         skb_pull(skb, skip_header_bytes);
 181         head_need = hdrlen + encaps_len + meshhdrlen - skb_headroom(skb);
 182 +       head_need += padsize;
 183
 184         /*
 185          * So we need to modify the skb header and hence need a copy of
 186 @@ -2734,6 +2737,9 @@ static struct sk_buff *ieee80211_build_h
 187                 memcpy(skb_push(skb, meshhdrlen), &mesh_hdr, meshhdrlen);
 188  #endif
 189
 190 +       if (padsize)
 191 +               memset(skb_push(skb, padsize), 0, padsize);
 192 +
 193         if (ieee80211_is_data_qos(fc)) {
 194                 __le16 *qos_control;
 195
 196 @@ -2909,6 +2915,9 @@ void ieee80211_check_fast_xmit(struct st
 197                 fc |= cpu_to_le16(IEEE80211_STYPE_QOS_DATA);
 198         }
 199
 200 +       /* Check aligned4 skb required */
 201 +       build.hdr_len += ieee80211_hdr_padsize(&local->hw, build.hdr_len);
 202 +
 203         /* We store the key here so there's no point in using rcu_dereference()
 204          * but that's fine because the code that changes the pointers will call
 205          * this function after doing so. For a single CPU that would be enough,
 206 @@ -3495,7 +3504,7 @@ begin:
 207
 208                 if (tx.key &&
 209                     (tx.key->conf.flags & IEEE80211_KEY_FLAG_GENERATE_IV))
 210 -                       pn_offs = ieee80211_hdrlen(hdr->frame_control);
 211 +                       pn_offs = tx.hdrlen;
 212
 213                 ieee80211_xmit_fast_finish(sta->sdata, sta, pn_offs,
 214                                            tx.key, skb);
 215 --- a/net/mac80211/util.c
 216 +++ b/net/mac80211/util.c
 217 @@ -1230,6 +1230,7 @@ void ieee80211_send_auth(struct ieee8021
 218                          u32 tx_flags)
 219  {
 220         struct ieee80211_local *local = sdata->local;
 221 +       struct ieee80211_hw *hw = &local->hw;
 222         struct sk_buff *skb;
 223         struct ieee80211_mgmt *mgmt;
 224         unsigned int hdrlen;
 225 @@ -1256,7 +1257,7 @@ void ieee80211_send_auth(struct ieee8021
 226                 skb_put_data(skb, extra, extra_len);
 227
 228         if (auth_alg == WLAN_AUTH_SHARED_KEY && transaction == 3) {
 229 -               hdrlen = ieee80211_hdrlen(mgmt->frame_control);
 230 +               hdrlen = ieee80211_padded_hdrlen(hw, mgmt->frame_control);
 231                 mgmt->frame_control |= cpu_to_le16(IEEE80211_FCTL_PROTECTED);
 232                 err = ieee80211_wep_encrypt(local, skb, hdrlen, key,
 233                                             key_len, key_idx);