1 From: Felix Fietkau <nbd@nbd.name>
2 Date: Sun, 17 Mar 2019 18:11:30 +0100
3 Subject: [PATCH] mac80211: optimize skb resizing
5 When forwarding unicast packets from ethernet to batman-adv over 802.11s
6 (with forwarding disabled), the typical required headroom to transmit
7 encrypted packets on mt76 is 32 (802.11) + 6 (802.11s) + 8 (CCMP) +
8 2 (padding) + 6 (LLC) + 18 (batman-adv) - 14 (old ethernet header) = 58 bytes.
10 On systems where NET_SKB_PAD is 64 this leads to a call to pskb_expand_head
11 for every packet, since mac80211 also tries to allocate 16 bytes status
12 headroom for radiotap headers.
14 This patch fixes these unnecessary reallocations by only requiring the extra
15 status headroom in ieee80211_tx_monitor()
16 If however a reallocation happens before that call, the status headroom gets
17 added there as well, in order to avoid double reallocation.
19 The patch also cleans up the code by moving the headroom calculation to
22 Signed-off-by: Felix Fietkau <nbd@nbd.name>
25 --- a/net/mac80211/ieee80211_i.h
26 +++ b/net/mac80211/ieee80211_i.h
27 @@ -1779,6 +1779,9 @@ int ieee80211_tx_control_port(struct wip
28 const u8 *dest, __be16 proto, bool unencrypted);
29 int ieee80211_probe_mesh_link(struct wiphy *wiphy, struct net_device *dev,
30 const u8 *buf, size_t len);
31 +int ieee80211_skb_resize(struct ieee80211_local *local,
32 + struct ieee80211_sub_if_data *sdata,
33 + struct sk_buff *skb, int hdrlen, int hdr_add);
36 void ieee80211_apply_htcap_overrides(struct ieee80211_sub_if_data *sdata,
37 --- a/net/mac80211/status.c
38 +++ b/net/mac80211/status.c
39 @@ -655,6 +655,11 @@ void ieee80211_tx_monitor(struct ieee802
40 struct net_device *prev_dev = NULL;
43 + if (ieee80211_skb_resize(local, NULL, skb, 0, 0)) {
48 /* send frame to monitor interfaces now */
49 rtap_len = ieee80211_tx_radiotap_len(info);
50 if (WARN_ON_ONCE(skb_headroom(skb) < rtap_len)) {
51 --- a/net/mac80211/tx.c
52 +++ b/net/mac80211/tx.c
53 @@ -1936,37 +1936,53 @@ static bool ieee80211_tx(struct ieee8021
56 /* device xmit handlers */
58 -static int ieee80211_skb_resize(struct ieee80211_sub_if_data *sdata,
59 - struct sk_buff *skb,
60 - int head_need, bool may_encrypt)
61 +int ieee80211_skb_resize(struct ieee80211_local *local,
62 + struct ieee80211_sub_if_data *sdata,
63 + struct sk_buff *skb, int hdr_len, int hdr_extra)
65 - struct ieee80211_local *local = sdata->local;
66 + struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
67 struct ieee80211_hdr *hdr;
70 + int head_need, head_max;
71 + int tail_need, tail_max;
72 + bool enc_tailroom = false;
74 - hdr = (struct ieee80211_hdr *) skb->data;
75 - enc_tailroom = may_encrypt &&
76 - (sdata->crypto_tx_tailroom_needed_cnt ||
77 - ieee80211_is_mgmt(hdr->frame_control));
78 + if (sdata && !hdr_len &&
79 + !(info->flags & IEEE80211_TX_INTFL_DONT_ENCRYPT)) {
80 + hdr = (struct ieee80211_hdr *) skb->data;
81 + enc_tailroom = (sdata->crypto_tx_tailroom_needed_cnt ||
82 + ieee80211_is_mgmt(hdr->frame_control));
83 + hdr_len += sdata->encrypt_headroom;
87 - tail_need = IEEE80211_ENCRYPT_TAILROOM;
88 - tail_need -= skb_tailroom(skb);
89 - tail_need = max_t(int, tail_need, 0);
90 + head_need = head_max = hdr_len;
91 + tail_need = tail_max = 0;
93 + head_need = head_max = local->tx_headroom;
95 + head_max += hdr_extra;
96 + head_max += max_t(int, local->tx_headroom,
97 + local->hw.extra_tx_headroom);
98 + head_need += local->hw.extra_tx_headroom;
100 + tail_max = IEEE80211_ENCRYPT_TAILROOM;
102 + tail_need = tail_max;
105 if (skb_cloned(skb) &&
106 (!ieee80211_hw_check(&local->hw, SUPPORTS_CLONED_SKBS) ||
107 !skb_clone_writable(skb, ETH_HLEN) || enc_tailroom))
108 I802_DEBUG_INC(local->tx_expand_skb_head_cloned);
109 - else if (head_need || tail_need)
110 + else if (head_need > skb_headroom(skb) ||
111 + tail_need > skb_tailroom(skb))
112 I802_DEBUG_INC(local->tx_expand_skb_head);
116 - if (pskb_expand_head(skb, head_need, tail_need, GFP_ATOMIC)) {
117 + head_max = max_t(int, 0, head_max - skb_headroom(skb));
118 + tail_max = max_t(int, 0, tail_max - skb_tailroom(skb));
120 + if (pskb_expand_head(skb, head_max, tail_max, GFP_ATOMIC)) {
121 wiphy_debug(local->hw.wiphy,
122 "failed to reallocate TX buffer\n");
124 @@ -1982,18 +1998,8 @@ void ieee80211_xmit(struct ieee80211_sub
125 struct ieee80211_local *local = sdata->local;
126 struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
127 struct ieee80211_hdr *hdr;
131 - may_encrypt = !(info->flags & IEEE80211_TX_INTFL_DONT_ENCRYPT);
133 - headroom = local->tx_headroom;
135 - headroom += sdata->encrypt_headroom;
136 - headroom -= skb_headroom(skb);
137 - headroom = max_t(int, 0, headroom);
139 - if (ieee80211_skb_resize(sdata, skb, headroom, may_encrypt)) {
140 + if (ieee80211_skb_resize(local, sdata, skb, 0, 0)) {
141 ieee80211_free_txskb(&local->hw, skb);
144 @@ -2774,29 +2780,13 @@ static struct sk_buff *ieee80211_build_h
147 skb_pull(skb, skip_header_bytes);
148 - head_need = hdrlen + encaps_len + meshhdrlen - skb_headroom(skb);
151 - * So we need to modify the skb header and hence need a copy of
152 - * that. The head_need variable above doesn't, so far, include
153 - * the needed header space that we don't need right away. If we
154 - * can, then we don't reallocate right now but only after the
155 - * frame arrives at the master device (if it does...)
157 - * If we cannot, however, then we will reallocate to include all
158 - * the ever needed space. Also, if we need to reallocate it anyway,
159 - * make it big enough for everything we may ever need.
161 + head_need = hdrlen + encaps_len + meshhdrlen;
163 - if (head_need > 0 || skb_cloned(skb)) {
164 - head_need += sdata->encrypt_headroom;
165 - head_need += local->tx_headroom;
166 - head_need = max_t(int, 0, head_need);
167 - if (ieee80211_skb_resize(sdata, skb, head_need, true)) {
168 - ieee80211_free_txskb(&local->hw, skb);
170 - return ERR_PTR(-ENOMEM);
172 + if (ieee80211_skb_resize(local, sdata, skb, head_need,
173 + sdata->encrypt_headroom)) {
174 + ieee80211_free_txskb(&local->hw, skb);
176 + return ERR_PTR(-ENOMEM);
180 @@ -3411,7 +3401,6 @@ static bool ieee80211_xmit_fast(struct i
181 struct ieee80211_local *local = sdata->local;
182 u16 ethertype = (skb->data[12] << 8) | skb->data[13];
183 int extra_head = fast_tx->hdr_len - (ETH_HLEN - 2);
184 - int hw_headroom = sdata->local->hw.extra_tx_headroom;
186 struct ieee80211_tx_info *info;
187 struct ieee80211_hdr *hdr = (void *)fast_tx->hdr;
188 @@ -3463,10 +3452,7 @@ static bool ieee80211_xmit_fast(struct i
189 * as the may-encrypt argument for the resize to not account for
190 * more room than we already have in 'extra_head'
192 - if (unlikely(ieee80211_skb_resize(sdata, skb,
193 - max_t(int, extra_head + hw_headroom -
194 - skb_headroom(skb), 0),
196 + if (unlikely(ieee80211_skb_resize(local, sdata, skb, extra_head, 0))) {