mac80211: fix a tx queue memory accounting error

Fixes rare hard to trigger tx hangs after some time

Signed-off-by: Felix Fietkau <nbd@nbd.name>
v19.07.3_mercusys_ac12_duma
Felix Fietkau 6 years ago
parent b7b14fd64e
commit 6011f7bcf0

@ -0,0 +1,66 @@
From: Felix Fietkau <nbd@nbd.name>
Date: Thu, 8 Mar 2018 21:00:56 +0100
Subject: [PATCH] mac80211: fix memory accounting with A-MSDU aggregation
fq uses skb->truesize for memory usage tracking. Increments/decrements
are done on enqueue/dequeue.
When A-MSDU aggregation is performed on tx side, the packet is
aggregated with the last packet in the queue belonging to the same flow.
There are multiple bugs here:
- The truesize field of the aggregated packet isn't updated, so memory
usage is underestimated
- fq->memory_usage isn't adjusted.
Because of the combination of both bugs, this only causes tx issues in
rare cases, mainly when the A-MSDU head needs to be reallocated.
Fix this by adjusting both truesize of the A-MSDU head and adding the
truesize delta to fq->memory_usage.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
---
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -3171,6 +3171,7 @@ static bool ieee80211_amsdu_aggregate(st
u8 max_subframes = sta->sta.max_amsdu_subframes;
int max_frags = local->hw.max_tx_fragments;
int max_amsdu_len = sta->sta.max_amsdu_len;
+ int orig_truesize;
__be16 len;
void *data;
bool ret = false;
@@ -3201,12 +3202,13 @@ static bool ieee80211_amsdu_aggregate(st
flow = fq_flow_classify(fq, tin, skb, fq_flow_get_default_func);
head = skb_peek_tail(&flow->queue);
if (!head)
- goto out;
+ goto unlock;
+ orig_truesize = head->truesize;
orig_len = head->len;
if (skb->len + head->len > max_amsdu_len)
- goto out;
+ goto unlock;
if (!ieee80211_amsdu_prepare_head(sdata, fast_tx, head))
goto out;
@@ -3241,6 +3243,7 @@ static bool ieee80211_amsdu_aggregate(st
head->len += skb->len;
head->data_len += skb->len;
+ head->truesize += skb->truesize;
*frag_tail = skb;
flow->backlog += head->len - orig_len;
@@ -3249,6 +3252,9 @@ static bool ieee80211_amsdu_aggregate(st
fq_recalc_backlog(fq, tin, flow);
out:
+ fq->memory_usage += head->truesize - orig_truesize;
+
+unlock:
spin_unlock_bh(&fq->lock);
return ret;
Loading…
Cancel
Save