A-MSDU reception not working?
Adrian Chadd
adrian at freebsd.org
Mon Jul 7 12:41:32 PDT 2014
.. why not disable AMSDU reception for now? Until these things are fixed?
-a
On 7 July 2014 12:26, Denton Gentry <denton.gentry at gmail.com> wrote:
> The initial results are not promising: a MacOS 802.11ac client gets
> between 0-2 Mbps with this change, where it was getting about 8 Mbps
> prior to this change and ~170 Mbps prior to the reordering fix. A pcap
> from the receiving system shows a very large number of out of order
> frames, likely due to TCP retransmission.
>
> An 802.11n MacBook gets very good throughput, only the 802.11ac
> MacBook shows the very poor result. I'm trying to figure out why.
>
>
> One specific note (and probably not related to the throughput): I
> believe ath10k_htt_rx_amsdu runs in the tasklet, which means it would
> need to use GFP_ATOMIC instead of GFP_KERNEL. Do I understand it
> correctly?
>
>
> On Mon, Jul 7, 2014 at 1:30 AM, Janusz Dziedzic
> <janusz.dziedzic at tieto.com> wrote:
>> On 6 July 2014 04:27, Adrian Chadd <adrian at freebsd.org> wrote:
>>> I think you may have to tell mac80211 that it's okay and not to drop the frames.
>>>
>>> The earlier atheros chips would just give you the AMSDU frames as
>>> deaggregated A-MPDU sub-frames - you'd just pass the A-MSDU up to
>>> net80211 and it'd ull it apart. But if the driver is doing it (or,
>>> well, the chip is doing it) then mac80211 needs to know not to drop
>>> those sub-frames.
>>>
>>> I wonder if you'll ever get notified before the complete A-MPDU has
>>> been received. That happens on previous chips. eg, you have an A-MPDU
>>> of 16 frames with 4 MSDUs in each MPDU. If you get notified and handle
>>> half of one MPDU before you hit EOL, the next notification you process
>>> will be the next MSDU in the same MPDU - and then you'll hit the
>>> reordering hilarity again.
>>>
>>> So hm, i will face the same issue in FreeBSD at some point, so I'd
>>> likely do what you're thinking of doing - pass up a chain of mbufs
>>> representing the current MPDU and treat the whole lot as the frame(s)
>>> to care about. Honestly though, I'm kind of wondering whether I should
>>> mostly just do what the Atheros reference driver does and treat it as
>>> ethernet payload frames (ie, it's already de-encapsulated and the
>>> reordering is already done) and just tack on the wifi header bit via
>>> another of the DMA rings.
>>>
>>> Ugh, I really should sit down and write the FreeBSD version of this.
>>>
>>>
>>>
>>> -a
>>>
>>> (I'm still having flashbacks from working on this firmware at QCA. Aiee.)
>>>
>>>
>>> On 5 July 2014 06:55, Denton Gentry <denton.gentry at gmail.com> wrote:
>>>> There are two issues in handling the dis-aggregated A-MSDU subframes
>>>> in ieee80211_sta_manage_reorder_buf:
>>>>
>>>> 1. The out-of-date check:
>>>>
>>>> /* frame with out of date sequence number */
>>>> if (ieee80211_sn_less(mpdu_seq_num, head_seq_num)) {
>>>> dev_kfree_skb(skb);
>>>> goto out;
>>>> }
>>>>
>>>> As all of the subframes carry the same sequence number, the first
>>>> subframe will be delivered and increment head_seq_num and then all
>>>> subsequent subframes will be discarded.
>>>>
>>>>
>>>> 2. The duplicates check a bit later in the routine:
>>>>
>>>> /* check if we already stored this frame */
>>>> if (tid_agg_rx->reorder_buf[index]) {
>>>> dev_kfree_skb(skb);
>>>> goto out;
>>>> }
>>>>
>>>> If there is enough packet loss that packets are queued in the reorder
>>>> buffer and not immediately released, then only the first subframe will
>>>> be stored. Subsequent subframes will be discarded as duplicates.
>>>>
>>>>
>>>> An 802.11ac MacBook is able to get about 170 Mbps with iperf prior to
>>>> the reordering buffer changes, and dropped to about 8 Mbps with the
>>>> reorder buffer. Hacking around the out-of-date sequence number check
>>>> to allow all subframes to egress restores it to 170 Mbps.
>>>>
>>>> In the area where I'm testing there isn't enough 5 GHz noise to make
>>>> the duplicates-check issue happen very often. By adding a printk I can
>>>> see that it does happen, but it doesn't impact the throughput and I
>>>> can't report the impact of fixing it.
>>>>
>>>> ----
>>>>
>>>> I do wonder if both of these are symptoms of dis-aggregating the
>>>> A-MSDU too early. mac80211 expects to be dealing with the whole MPDU
>>>> at a time, and the ath10k A-MSDU case is sending it subframes instead.
>>>> Trying to make the ath10k code send up all of the subframes as a chain
>>>> of skbs didn't immediately work, but I do wonder if that would better
>>>> match the mac80211 expectations.
>>>>
>>>>
>>>>
>>>> On Fri, Jul 4, 2014 at 11:58 AM, Denton Gentry <denton.gentry at gmail.com> wrote:
>>>>> Yes, after some more testing it does look like an unfortunate
>>>>> interaction between the reorder buffer and A-MSDU. The disaggregated
>>>>> subframes all carry the same sequence number. The first subframe is
>>>>> released from the reorder buffer and increments the head_seq_num.
>>>>> Subsequent subframes are all discarded as being out of date.
>>>>>
>>>>> [ 308.514021] ieee80211_sta_manage_reorder_buf: out of date seq=0xb05
>>>>> head=0xb06
>>>>> [ 308.520577] ieee80211_sta_manage_reorder_buf: out of date seq=0xb0a
>>>>> head=0xb0b
>>>>> [ 308.527198] ieee80211_sta_manage_reorder_buf: out of date seq=0xb0f
>>>>> head=0xb10
>>>>> [ 308.533857] ieee80211_sta_manage_reorder_buf: out of date seq=0xb14
>>>>> head=0xb15
>>>>> [ 308.540480] ieee80211_sta_manage_reorder_buf: out of date seq=0xb19
>>>>> head=0xb1a
>>>>> [ 308.547730] ieee80211_sta_manage_reorder_buf: out of date seq=0xb1e
>>>>> head=0xb1f
>>>>> [ 308.554069] ieee80211_sta_manage_reorder_buf: out of date seq=0xb23
>>>>> head=0xb24
>>>>>
>>>>>
>>>>> On Wed, Jul 2, 2014 at 9:49 AM, Michal Kazior <michal.kazior at tieto.com> wrote:
>>>>>> On 30 June 2014 22:15, Denton Gentry <denton.gentry at gmail.com> wrote:
>>>>>>> In iperf tests using a MacBook STA bridging through an ath10k AP to an
>>>>>>> Ethernet server, I'm noticing very selective packet loss. The second
>>>>>>> and subsequent frames in an A-MSDU packet appear to be dropped.
>>>>>>>
>>>>>>> The AP sets the A-MSDU size to 3839 bytes, and the MacBook frequently
>>>>>>> sends A-MSDU packets containing two TCP frames. So far as I can tell,
>>>>>>> the first TCP frame from an A-MSDU aggregate is delivered and the
>>>>>>> second is consistently lost. The MacBook generally retransmits the
>>>>>>> lost frame as a singleton with no aggregation, and the retransmitted
>>>>>>> frame makes it through.
>>>>>>>
>>>>>>> This became more noticeable after the reordering fixes in
>>>>>>> http://lists.infradead.org/pipermail/ath10k/2014-June/002552.html
>>>>>>>
>>>>>>> I see this A-MSDU packet loss behavior both with and without the
>>>>>>> reordering fixes, the first packet in an A-MSDU is delivered while the
>>>>>>> second is dropped. However, *without* the reordering fixes (and
>>>>>>> therefore with packets delivered out of order) the MacBook sends
>>>>>>> relatively few A-MSDU frames. *With* the reordering fixes, so all
>>>>>>> packets are delivered in order, the MacBook keeps sending A-MSDU and
>>>>>>> therefore has to deal with more packet loss. I suspect it is an
>>>>>>> interaction with the MacOS TCP congestion window which I'm likely
>>>>>>> never going to fully understand, its stuck in a region of the
>>>>>>> congestion window where the Wifi driver keeps choosing to using
>>>>>>> A-MSDU.
>>>>>>
>>>>>> I was actually worried about A-MSDU once A-MPDU re-ordering issue was raised.
>>>>>>
>>>>>> ath10k fw reports A-MSDU subframes separately. To avoid memory
>>>>>> copying/allocation overhead each subframe is reported as a singly
>>>>>> A-MSDU MSDU to mac80211 with an extra rx_flag AMSDU_MORE. Perhaps
>>>>>> A-MPDU reordering intereferes with A-MSDU now?
>>>>>>
>> Denton could you try attached patch: report amsdu in one big frame.
>> If help, we can add amsdu skb list support to mac80211/cfg80211 - to
>> improve performance and reduce memcpy, while currently we have skb
>> frames, put them in one big skb and next cfg80211 split them again
>> into msdus and report to stack.
>>
>> BR
>> Janusz
>
> _______________________________________________
> ath10k mailing list
> ath10k at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/ath10k
More information about the ath10k
mailing list