From: Johannes Berg <[email protected]>
The description for buf_size was misleading and
just said you couldn't TX larger aggregates, but
of course you can't TX aggregates in a way that
would exceed the window either, which is possible
even if the aggregates are shorter than that.
Expand the description, thanks to Emmanuel for
explaining this to me.
Cc: Emmanuel Grumbach <[email protected]>
Signed-off-by: Johannes Berg <[email protected]>
---
include/net/mac80211.h | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
--- a/include/net/mac80211.h 2011-03-16 10:30:26.000000000 +0100
+++ b/include/net/mac80211.h 2011-03-16 10:37:49.000000000 +0100
@@ -1753,8 +1753,17 @@ enum ieee80211_ampdu_mlme_action {
* that TX/RX_STOP can pass NULL for this parameter.
* The @buf_size parameter is only valid when the action is set to
* %IEEE80211_AMPDU_TX_OPERATIONAL and indicates the peer's reorder
- * buffer size (number of subframes) for this session -- aggregates
- * containing more subframes than this may not be transmitted to the peer.
+ * buffer size (number of subframes) for this session -- the driver
+ * may neither send aggregates containing more subframes than this
+ * nor send aggregates in a way that lost frames would exceed the
+ * buffer size. If just limiting the aggregate size, this would be
+ * possible with a buf_size of 8:
+ * - TX: 1......8
+ * - RX: 2.....8 (lost frame #1)
+ * - TX: 9...1...
+ * which is invalid since #1 was now re-transmitted well past the
+ * buffer size of 8. The correct way to retransmit #1 would be:
+ * - TX: 19......
* Returns a negative error code on failure.
* The callback can sleep.
*
On Wed, 2011-03-16 at 02:47 -0700, Daniel Halperin wrote:
> <dan starts here>
>
> > + * possible with a buf_size of 8:
> > + * - TX: 1.....7
> > + * - RX: 2.....7 (lost frame #1),
> > + * - TX: 89...1...
> > + * which is invalid since #1 was now re-transmitted well past the
> > + * buffer size of 7. The correct way to retransmit #1 would be:
> > + * - TX: 1 or
> > + * - TX: 18
> > + * though the standard allows:
> > + * - TX: 81
> > + * The following might happen, but will violate buffering assumptions as above if frame 1 is lost again
> > + * - TX: 189.....
Oh, oops, yes you're right. Will respin the patch.
johannes
That last line should really be:
>> + * ? ? Note that the following is also illegal because it will violate buffering assumptions if frame 1 is lost again
>> + * ? ? ?- TX: ? ? ? ? 189.....
>
> Dan
>
On Wed, Mar 16, 2011 at 11:51, Daniel Halperin
<[email protected]> wrote:
> That last line should really be:
>
>>> + * ? ? Note that the following is also illegal because it will violate buffering assumptions if frame 1 is lost again
agreed - the Txer cannot send any frame with seqNum greater than SSN
from last BA reponse + window size. In this case, the SSN of last BA
is 0, so you cannot send 8 and any further frame until you get the BA
response for 1.
>>> + * ? ? ?- TX: ? ? ? ? 189.....
>>
>> Dan
>>
>
Hmm, this seems to not quite convey some complexities. See some inlines below.
On Wed, Mar 16, 2011 at 2:39 AM, Johannes Berg
<[email protected]> wrote:
> From: Johannes Berg <[email protected]>
>
> The description for buf_size was misleading and
> just said you couldn't TX larger aggregates, but
> of course you can't TX aggregates in a way that
> would exceed the window either, which is possible
> even if the aggregates are shorter than that.
>
> Expand the description, thanks to Emmanuel for
> explaining this to me.
>
> Cc: Emmanuel Grumbach <[email protected]>
> Signed-off-by: Johannes Berg <[email protected]>
> ---
> ?include/net/mac80211.h | ? 13 +++++++++++--
> ?1 file changed, 11 insertions(+), 2 deletions(-)
>
> --- a/include/net/mac80211.h ? ?2011-03-16 10:30:26.000000000 +0100
> +++ b/include/net/mac80211.h ? ?2011-03-16 10:37:49.000000000 +0100
> @@ -1753,8 +1753,17 @@ enum ieee80211_ampdu_mlme_action {
> ?* ? ? that TX/RX_STOP can pass NULL for this parameter.
> ?* ? ? The @buf_size parameter is only valid when the action is set to
> ?* ? ? %IEEE80211_AMPDU_TX_OPERATIONAL and indicates the peer's reorder
> - * ? ? buffer size (number of subframes) for this session -- aggregates
> - * ? ? containing more subframes than this may not be transmitted to the peer.
> + * ? ? buffer size (number of subframes) for this session -- the driver
> + * ? ? may neither send aggregates containing more subframes than this
> + * ? ? nor send aggregates in a way that lost frames would exceed the
> + * ? ? buffer size. If just limiting the aggregate size, this would be
<dan starts here>
> + * ? ? possible with a buf_size of 8:
> + * ? ? ?- TX: 1.....7
> + * ? ? ?- RX: 2.....7 (lost frame #1),
> + * ? ? ?- TX: ? ? ? ? 89...1...
> + * ? ? which is invalid since #1 was now re-transmitted well past the
> + * ? ? buffer size of 7. The correct way to retransmit #1 would be:
> + * ? ? ?- TX: ? ? ? ? 1 or
> + * ? ? ?- TX: ? ? ? ? 18
> + * ? ? though the standard allows:
> + * ? ? ?- TX: ? ? ? ? 81
> + * ? ? The following might happen, but will violate buffering assumptions as above if frame 1 is lost again
> + * ? ? ?- TX: ? ? ? ? 189.....
Dan