Patchwork [v3] iwl3945: better skb management in rx path

login
register
mail settings
Submitter Eric Dumazet
Date June 28, 2013, 3:05 p.m.
Message ID <1372431906.3301.283.camel@edumazet-glaptop>
Download mbox | patch
Permalink /patch/255415/
State Not Applicable
Delegated to: David Miller
Headers show

Comments

Eric Dumazet - June 28, 2013, 3:05 p.m.
From: Eric Dumazet <edumazet@google.com>

Steinar reported reallocations of skb->head with IPv6, leading to
a warning in skb_try_coalesce()

It turns out iwl3945 has several problems :

1) skb->truesize is underestimated. 
   We really consume PAGE_SIZE bytes for a fragment,
   not the frame length.
2) 128 bytes of initial headroom is a bit low and forces reallocations.
3) We can avoid consuming a full page for small enough frames.

Reported-by: Steinar H. Gunderson <sesse@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paul Stewart <pstew@google.com>
---
v3: use regular memcpy(skb_put(...),...)
v2: SMALL_PACKET_SIZE define

 drivers/net/wireless/iwlegacy/3945.c |   31 +++++++++++++++----------
 1 file changed, 19 insertions(+), 12 deletions(-)



--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Steinar H. Gunderson - June 28, 2013, 11:05 p.m.
2013/6/28 Eric Dumazet <eric.dumazet@gmail.com>:
> Steinar reported reallocations of skb->head with IPv6, leading to
> a warning in skb_try_coalesce()
>
> It turns out iwl3945 has several problems :
>
> 1) skb->truesize is underestimated.
>    We really consume PAGE_SIZE bytes for a fragment,
>    not the frame length.
> 2) 128 bytes of initial headroom is a bit low and forces reallocations.
> 3) We can avoid consuming a full page for small enough frames.

I'm running 3.9.8 plus this patch (hand-applied since GMail mangled
it…), and so far, it seems to work fine. Well, network-manager still
doesn't like Cisco's band-select feature, so from time to time I get
network freezes while it's trying to roam from 5 GHz to 2.4 GHz, but
that's hardly iwl3945's fault.

/* Steinar */
Stanislaw Gruszka - July 1, 2013, 12:15 p.m.
On Fri, Jun 28, 2013 at 08:05:06AM -0700, Eric Dumazet wrote:
> From: Eric Dumazet <edumazet@google.com>
> 
> Steinar reported reallocations of skb->head with IPv6, leading to
> a warning in skb_try_coalesce()
> 
> It turns out iwl3945 has several problems :
> 
> 1) skb->truesize is underestimated. 
>    We really consume PAGE_SIZE bytes for a fragment,
>    not the frame length.
> 2) 128 bytes of initial headroom is a bit low and forces reallocations.
> 3) We can avoid consuming a full page for small enough frames.
> 
> Reported-by: Steinar H. Gunderson <sesse@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Paul Stewart <pstew@google.com>

Acked-by: Stanislaw Gruszka <sgruszka@redhat.com>

FYI, I'll post 4965 version soon.


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/drivers/net/wireless/iwlegacy/3945.c b/drivers/net/wireless/iwlegacy/3945.c
index c092033..f09e257 100644
--- a/drivers/net/wireless/iwlegacy/3945.c
+++ b/drivers/net/wireless/iwlegacy/3945.c
@@ -475,6 +475,8 @@  il3945_is_network_packet(struct il_priv *il, struct ieee80211_hdr *header)
 	}
 }
 
+#define SMALL_PACKET_SIZE 256
+
 static void
 il3945_pass_packet_to_mac80211(struct il_priv *il, struct il_rx_buf *rxb,
 			       struct ieee80211_rx_status *stats)
@@ -483,14 +485,13 @@  il3945_pass_packet_to_mac80211(struct il_priv *il, struct il_rx_buf *rxb,
 	struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)IL_RX_DATA(pkt);
 	struct il3945_rx_frame_hdr *rx_hdr = IL_RX_HDR(pkt);
 	struct il3945_rx_frame_end *rx_end = IL_RX_END(pkt);
-	u16 len = le16_to_cpu(rx_hdr->len);
+	u32 len = le16_to_cpu(rx_hdr->len);
 	struct sk_buff *skb;
 	__le16 fc = hdr->frame_control;
+	u32 fraglen = PAGE_SIZE << il->hw_params.rx_page_order;
 
 	/* We received data from the HW, so stop the watchdog */
-	if (unlikely
-	    (len + IL39_RX_FRAME_SIZE >
-	     PAGE_SIZE << il->hw_params.rx_page_order)) {
+	if (unlikely(len + IL39_RX_FRAME_SIZE > fraglen)) {
 		D_DROP("Corruption detected!\n");
 		return;
 	}
@@ -506,26 +507,32 @@  il3945_pass_packet_to_mac80211(struct il_priv *il, struct il_rx_buf *rxb,
 		D_INFO("Woke queues - frame received on passive channel\n");
 	}
 
-	skb = dev_alloc_skb(128);
+	skb = dev_alloc_skb(SMALL_PACKET_SIZE);
 	if (!skb) {
 		IL_ERR("dev_alloc_skb failed\n");
 		return;
 	}
 
 	if (!il3945_mod_params.sw_crypto)
-		il_set_decrypted_flag(il, (struct ieee80211_hdr *)rxb_addr(rxb),
+		il_set_decrypted_flag(il, (struct ieee80211_hdr *)pkt,
 				      le32_to_cpu(rx_end->status), stats);
 
-	skb_add_rx_frag(skb, 0, rxb->page,
-			(void *)rx_hdr->payload - (void *)pkt, len,
-			len);
-
+	/* If frame is small enough to fit into skb->head, copy it
+	 * and do not consume a full page
+	 */
+	if (len <= SMALL_PACKET_SIZE) {
+		memcpy(skb_put(skb, len), rx_hdr->payload, len);
+	} else {
+		skb_add_rx_frag(skb, 0, rxb->page,
+				(void *)rx_hdr->payload - (void *)pkt, len,
+				fraglen);
+		il->alloc_rxb_page--;
+		rxb->page = NULL;
+	}
 	il_update_stats(il, false, fc, len);
 	memcpy(IEEE80211_SKB_RXCB(skb), stats, sizeof(*stats));
 
 	ieee80211_rx(il->hw, skb);
-	il->alloc_rxb_page--;
-	rxb->page = NULL;
 }
 
 #define IL_DELAY_NEXT_SCAN_AFTER_ASSOC (HZ*6)