Re: [PATCH v2] mac80211: do not aggregate frames if max_frags is set to one

[Lorenzo Bianconi <lorenzo.bianconi@redhat.com>]

> On Thu, 2018-08-30 at 10:31 +0200, Lorenzo Bianconi wrote:
> 
> > Reviewing the code I guess it is not necessary since pskb_expand_head routine
> > does not modify head->len (or skb->len). 
> 
> True.
> 
> > Packet len (if we consider padding) is only modified in:
> > 
> > memset(skb_push(skb, pad), 0, pad);
> > 
> > and if we hit that point, we will account new skb->len in flow backlog. Do you
> > agree?
> 
> Right, but that's the *pad*. I was thinking about the header conversion.
> 
> Let's say you decided to add the second frame to the A-MSDU, at which
> point the first one isn't really an A-MSDU yet. So we get to:
> 
>         if (!ieee80211_amsdu_prepare_head(sdata, fast_tx, head))
> 
> which changes the header of "head" to be 14 bytes longer:
> 
> 	skb_push(skb, sizeof(*amsdu_hdr));
> 
> But now let's say we get a failure here when reallocating the second
> subframe:
> 
>         if (!ieee80211_amsdu_realloc_pad(local, skb, sizeof(rfc1042_header) +
>                                                      2 + pad))
>                 goto out;
> 
> Now we have changed "head", which is on the FQ, but we haven't changed
> the FQ accounting. So I *think* we still need this:
> 
> --- a/net/mac80211/tx.c
> +++ b/net/mac80211/tx.c
> @@ -3239,7 +3239,7 @@ static bool ieee80211_amsdu_aggregate(struct ieee80211_sub_if_data *sdata,
>  
>  	if (!ieee80211_amsdu_realloc_pad(local, skb, sizeof(rfc1042_header) +
>  						     2 + pad))
> -		goto out;
> +		goto out_recalc;
>  
>  	ret = true;
>  	data = skb_push(skb, ETH_ALEN + 2);
> @@ -3256,11 +3256,13 @@ static bool ieee80211_amsdu_aggregate(struct ieee80211_sub_if_data *sdata,
>  	head->data_len += skb->len;
>  	*frag_tail = skb;
>  
> -	flow->backlog += head->len - orig_len;
> -	tin->backlog_bytes += head->len - orig_len;
> -
> -	fq_recalc_backlog(fq, tin, flow);
> +out_recalc:
> +	if (head->len != orig_len) {
> +		flow->backlog += head->len - orig_len;
> +		tin->backlog_bytes += head->len - orig_len;
>  
> +		fq_recalc_backlog(fq, tin, flow);
> +	}
>  out:
>  	spin_unlock_bh(&fq->lock);
>  

ack, I agree. Do you want I send a patch to fix it?

> 
> 
> > Looking at the code maybe I spotted another issue, I guess there is an
> > off-by-one issue in 'n' estimation since it does not take into account
> > the first frame. We hit the line:
> > 
> > while (*frag_tail) {
> > }
> > 
> > starting from the second subframe, but if the head does not have packet in the
> > fraglist we will end up having n = 1, while it is actually the second frame.
> 
> Hmm, not sure I follow? "head" is the A-MSDU, containing the A-MSDU
> header and the first subframe in skb->data (and/or frags), with the
> subframes 2..N in the fraglist.
> 
> So I think this is right?

yep, correct. But when we are analyzing the second subframe what is the correct value for 'n'?
1 or 2? At the moment I guess it is set to 1 if frag_tail is NULL for head.

Regards,
Lorenzo

> 
> johannes
>