2013-09-23 13:35:16

by Michal Kazior

Subject: [PATCH] mac80211: support reporting A-MSDU subframes individually

Some devices may not be able to report A-MSDUs in
single buffers. Drivers for such devices were
forced to re-assemble A-MSDUs which would then
be eventually disassembled by mac80211. This could
lead to CPU cache thrashing and poor performance.

Since A-MSDU has a single sequence number all
subframes share it. This was in conflict with
retransmission/duplication recovery
(IEEE802.11-2012: 9.3.2.10).

Patch introduces a new flag that is meant to be
set for all individually reported A-MSDU subframes
except the last one. This ensures the
last_seq_ctrl is updated after the last subframe
is processed. If an A-MSDU is actually a duplicate
transmission all reported subframes will be
properly discarded.

Signed-off-by: Michal Kazior <[email protected]>
---
There were no comments on the RFC so far so I'm
posting this as a patch.

The first driver that will use this is ath10k.

include/net/mac80211.h | 10 ++++++++++
net/mac80211/rx.c | 2 +-
2 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/include/net/mac80211.h b/include/net/mac80211.h
index 06ffae8..f8f049d 100644
--- a/include/net/mac80211.h
+++ b/include/net/mac80211.h
@@ -829,6 +829,15 @@ ieee80211_tx_info_clear_status(struct ieee80211_tx_info *info)
* @RX_FLAG_STBC_MASK: STBC 2 bit bitmask. 1 - Nss=1, 2 - Nss=2, 3 - Nss=3
* @RX_FLAG_10MHZ: 10 MHz (half channel) was used
* @RX_FLAG_5MHZ: 5 MHz (quarter channel) was used
+ * @RX_FLAG_AMSDU_MORE: Some drivers may prefer to report separate A-MSDU
+ * subframes instead of a one huge frame for performance reasons.
+ * All, but the last MSDU from an A-MSDU should have this flag set. E.g.
+ * if an A-MSDU has 3 frames, the first 2 must have the flag set, while
+ * the 3rd (last) one must not have this flag set. The flag is used to
+ * deal with retransmission/duplication recovery properly since A-MSDU
+ * subframes share the same sequence number. Reported subframes can be
+ * either regular MSDU or singly A-MSDUs. Subframes must not be
+ * interleaved with other frames.
*/
enum mac80211_rx_flags {
RX_FLAG_MMIC_ERROR = BIT(0),
@@ -859,6 +868,7 @@ enum mac80211_rx_flags {
RX_FLAG_STBC_MASK = BIT(26) | BIT(27),
RX_FLAG_10MHZ = BIT(28),
RX_FLAG_5MHZ = BIT(29),
+ RX_FLAG_AMSDU_MORE = BIT(30),
};

#define RX_FLAG_STBC_SHIFT 26
diff --git a/net/mac80211/rx.c b/net/mac80211/rx.c
index 8e908e1..43bee02 100644
--- a/net/mac80211/rx.c
+++ b/net/mac80211/rx.c
@@ -995,7 +995,7 @@ ieee80211_rx_h_check(struct ieee80211_rx_data *rx)
rx->sta->num_duplicates++;
}
return RX_DROP_UNUSABLE;
- } else
+ } else if (!(status->flag & RX_FLAG_AMSDU_MORE))
rx->sta->last_seq_ctrl[rx->seqno_idx] = hdr->seq_ctrl;
}

--
1.7.9.5

2013-10-01 10:22:35

by Johannes Berg

Subject: Re: [PATCH] mac80211: support reporting A-MSDU subframes individually

On Mon, 2013-09-23 at 15:34 +0200, Michal Kazior wrote:
> Some devices may not be able to report A-MSDUs in
> single buffers. Drivers for such devices were
> forced to re-assemble A-MSDUs which would then
> be eventually disassembled by mac80211. This could
> lead to CPU cache thrashing and poor performance.
>
> Since A-MSDU has a single sequence number all
> subframes share it. This was in conflict with
> retransmission/duplication recovery
> (IEEE802.11-2012: 9.3.2.10).
>
> Patch introduces a new flag that is meant to be
> set for all individually reported A-MSDU subframes
> except the last one. This ensures the
> last_seq_ctrl is updated after the last subframe
> is processed. If an A-MSDU is actually a duplicate
> transmission all reported subframes will be
> properly discarded.

Applied.

johannes