Re: [PATCH 2/4] mac80211: defer txqs removal from rbtree

Yibo Zhao <yiboz@xxxxxxxxxxxxxx> · Sat, 21 Sep 2019 18:49:20 +0800

On 2019-09-20 17:15, Toke Høiland-Jørgensen wrote:
Yibo Zhao <yiboz@xxxxxxxxxxxxxx> writes:

On 2019-09-19 18:37, Toke Høiland-Jørgensen wrote:
Yibo Zhao <yiboz@xxxxxxxxxxxxxx> writes:

On 2019-09-18 19:23, Toke Høiland-Jørgensen wrote:
Yibo Zhao <yiboz@xxxxxxxxxxxxxx> writes:

On 2019-09-18 05:10, Toke Høiland-Jørgensen wrote:
Yibo Zhao <yiboz@xxxxxxxxxxxxxx> writes:

In a loop txqs dequeue scenario, if the first txq in the rbtree
gets
removed from rbtree immediately in the ieee80211_return_txq(), 
the
loop will break soon in the ieee80211_next_txq() due to
schedule_pos
not leading to the second txq in the rbtree. Thus, defering the
removal right before the end of this schedule round.

Co-developed-by: Yibo Zhao <yiboz@xxxxxxxxxxxxxx>
Signed-off-by: Yibo Zhao <yiboz@xxxxxxxxxxxxxx>
Signed-off-by: Toke Høiland-Jørgensen <toke@xxxxxxx>

I didn't write this patch, so please don't use my sign-off. I'll
add
ack or review tags as appropriate in reply; but a few comments
first:

---
 include/net/mac80211.h     | 16 ++++++++++--
 net/mac80211/ieee80211_i.h |  3 +++
 net/mac80211/main.c        |  6 +++++
 net/mac80211/tx.c          | 63
+++++++++++++++++++++++++++++++++++++++++++---
 4 files changed, 83 insertions(+), 5 deletions(-)

diff --git a/include/net/mac80211.h b/include/net/mac80211.h
index ac2ed8e..ba5a345 100644
--- a/include/net/mac80211.h
+++ b/include/net/mac80211.h
@@ -925,6 +925,8 @@ struct ieee80211_tx_rate {

 #define IEEE80211_MAX_TX_RETRY		31

+#define IEEE80211_AIRTIME_TXQ_RM_CHK_INTV_IN_MS 100
+
 static inline void ieee80211_rate_set_vht(struct
ieee80211_tx_rate
*rate,
 					  u8 mcs, u8 nss)
 {
@@ -6232,7 +6234,8 @@ struct sk_buff 
*ieee80211_tx_dequeue(struct
ieee80211_hw *hw,
  * @ac: AC number to return packets from.
  *
  * Should only be called between calls to
ieee80211_txq_schedule_start()
- * and ieee80211_txq_schedule_end().
+ * and ieee80211_txq_schedule_end(). If the txq is empty, it 
will
be
added
+ * to a remove list and get removed later.
  * Returns the next txq if successful, %NULL if no queue is
eligible.
If a txq
  * is returned, it should be returned with 
ieee80211_return_txq()
after the
  * driver has finished scheduling it.
@@ -6268,7 +6271,8 @@ void ieee80211_txq_schedule_start(struct
ieee80211_hw *hw, u8 ac)
  * @hw: pointer as obtained from ieee80211_alloc_hw()
  * @ac: AC number to acquire locks for
  *
- * Release locks previously acquired by
ieee80211_txq_schedule_end().
+ * Release locks previously acquired by
ieee80211_txq_schedule_end().
Check
+ * and remove the empty txq from rb-tree.
  */
 void ieee80211_txq_schedule_end(struct ieee80211_hw *hw, u8 ac)
 	__releases(txq_lock);
@@ -6287,6 +6291,14 @@ void ieee80211_schedule_txq(struct
ieee80211_hw
*hw, struct ieee80211_txq *txq)
 	__acquires(txq_lock) __releases(txq_lock);

 /**
+ * ieee80211_txqs_check - Check txqs waiting for removal
+ *
+ * @tmr: pointer as obtained from local
+ *
+ */
+void ieee80211_txqs_check(struct timer_list *tmr);
+
+/**
  * ieee80211_txq_may_transmit - check whether TXQ is allowed to
transmit
  *
  * This function is used to check whether given txq is allowed 
to
transmit by
diff --git a/net/mac80211/ieee80211_i.h
b/net/mac80211/ieee80211_i.h
index a4556f9..49aa143e 100644
--- a/net/mac80211/ieee80211_i.h
+++ b/net/mac80211/ieee80211_i.h
@@ -847,6 +847,7 @@ struct txq_info {
 	struct codel_stats cstats;
 	struct sk_buff_head frags;
 	struct rb_node schedule_order;
+	struct list_head candidate;
 	unsigned long flags;

 	/* keep last! */
@@ -1145,6 +1146,8 @@ struct ieee80211_local {
 	u64 airtime_v_t[IEEE80211_NUM_ACS];
 	u64 airtime_weight_sum[IEEE80211_NUM_ACS];

+	struct list_head remove_list[IEEE80211_NUM_ACS];
+	struct timer_list remove_timer;
 	u16 airtime_flags;

 	const struct ieee80211_ops *ops;
diff --git a/net/mac80211/main.c b/net/mac80211/main.c
index e9ffa8e..78fe24a 100644
--- a/net/mac80211/main.c
+++ b/net/mac80211/main.c
@@ -667,10 +667,15 @@ struct ieee80211_hw
*ieee80211_alloc_hw_nm(size_t priv_data_len,

 	for (i = 0; i < IEEE80211_NUM_ACS; i++) {
 		local->active_txqs[i] = RB_ROOT_CACHED;
+		INIT_LIST_HEAD(&local->remove_list[i]);
 		spin_lock_init(&local->active_txq_lock[i]);
 	}
 	local->airtime_flags = AIRTIME_USE_TX | AIRTIME_USE_RX;

+	timer_setup(&local->remove_timer, ieee80211_txqs_check, 0);
+	mod_timer(&local->remove_timer,
+		  jiffies +
msecs_to_jiffies(IEEE80211_AIRTIME_TXQ_RM_CHK_INTV_IN_MS));
+
 	INIT_LIST_HEAD(&local->chanctx_list);
 	mutex_init(&local->chanctx_mtx);

@@ -1305,6 +1310,7 @@ void ieee80211_unregister_hw(struct
ieee80211_hw
*hw)
 	tasklet_kill(&local->tx_pending_tasklet);
 	tasklet_kill(&local->tasklet);

+	del_timer_sync(&local->remove_timer);
 #ifdef CONFIG_INET
 	unregister_inetaddr_notifier(&local->ifa_notifier);
 #endif
diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
index d00baaa..42ca010 100644
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -1450,6 +1450,7 @@ void ieee80211_txq_init(struct
ieee80211_sub_if_data *sdata,
 	codel_stats_init(&txqi->cstats);
 	__skb_queue_head_init(&txqi->frags);
 	RB_CLEAR_NODE(&txqi->schedule_order);
+	INIT_LIST_HEAD(&txqi->candidate);

 	txqi->txq.vif = &sdata->vif;

@@ -3724,6 +3725,9 @@ void ieee80211_schedule_txq(struct
ieee80211_hw
*hw,

 	spin_lock_bh(&local->active_txq_lock[ac]);

+	if (!list_empty(&txqi->candidate))
+		list_del_init(&txqi->candidate);
+
 	if (!RB_EMPTY_NODE(&txqi->schedule_order))
 		goto out;

@@ -3783,6 +3787,20 @@ static void
__ieee80211_unschedule_txq(struct
ieee80211_hw *hw,
 	RB_CLEAR_NODE(&txqi->schedule_order);
 }

+void ieee80211_remove_txq(struct ieee80211_hw *hw,
+			  struct ieee80211_txq *txq)
+{
+	struct ieee80211_local *local = hw_to_local(hw);
+	struct txq_info *txqi = to_txq_info(txq);
+
+	lockdep_assert_held(&local->active_txq_lock[txq->ac]);
+
+	if (!RB_EMPTY_NODE(&txqi->schedule_order)) {
+		__ieee80211_unschedule_txq(hw, txq);
+		list_del_init(&txqi->candidate);
+	}
+}
+
 void ieee80211_unschedule_txq(struct ieee80211_hw *hw,
 			      struct ieee80211_txq *txq)
 	__acquires(txq_lock) __releases(txq_lock)
@@ -3790,7 +3808,7 @@ void ieee80211_unschedule_txq(struct
ieee80211_hw *hw,
 	struct ieee80211_local *local = hw_to_local(hw);

 	spin_lock_bh(&local->active_txq_lock[txq->ac]);
-	__ieee80211_unschedule_txq(hw, txq);
+	ieee80211_remove_txq(hw, txq);
 	spin_unlock_bh(&local->active_txq_lock[txq->ac]);
 }

@@ -3803,11 +3821,48 @@ void ieee80211_return_txq(struct
ieee80211_hw
*hw,
 	lockdep_assert_held(&local->active_txq_lock[txq->ac]);

 	if (!RB_EMPTY_NODE(&txqi->schedule_order) &&
-	    (skb_queue_empty(&txqi->frags) &&
!txqi->tin.backlog_packets))
-		__ieee80211_unschedule_txq(hw, txq);
+		!txq_has_queue(&txqi->txq) &&
+		list_empty(&txqi->candidate))
+		list_add_tail(&txqi->candidate, 
&local->remove_list[txq->ac]);
+
 }
 EXPORT_SYMBOL(ieee80211_return_txq);

+void __ieee80211_check_txqs(struct ieee80211_local *local, int
ac)
+{
+	struct txq_info *iter, *tmp;
+	struct sta_info *sta;
+
+	lockdep_assert_held(&local->active_txq_lock[ac]);
+
+	list_for_each_entry_safe(iter, tmp, &local->remove_list[ac],
+				 candidate) {
+		sta = container_of(iter->txq.sta, struct sta_info, sta);
+
+		if (txq_has_queue(&iter->txq))
+			list_del_init(&iter->candidate);
+		else
+			ieee80211_remove_txq(&local->hw, &iter->txq);
+	}
+}
+
+void ieee80211_txqs_check(struct timer_list *t)
+{
+	struct ieee80211_local *local = from_timer(local, t,
remove_timer);
+	struct txq_info *iter, *tmp;
+	struct sta_info *sta;
+	int ac;
+
+	for (ac = 0; ac < IEEE80211_NUM_ACS; ac++) {
+		spin_lock_bh(&local->active_txq_lock[ac]);
+		__ieee80211_check_txqs(local, ac);
+		spin_unlock_bh(&local->active_txq_lock[ac]);
+	}
+
+	mod_timer(&local->remove_timer,
+		  jiffies +
msecs_to_jiffies(IEEE80211_AIRTIME_TXQ_RM_CHK_INTV_IN_MS));
+}

I'll ask the same as I did last time (where you told me to hold 
off
until this round):

Why do you need the timer and the periodic check? If TXQs are 
added
to
the remove list during the scheduling run, and
__ieee80211_check_txqs()
is run from schedule_end(), isn't that sufficient to clear the
list?
Is it possible that a txq is not added to the remove list but then
packets in it are dropped by fq_codel algo? Like the station
disconnects
without any notification.

Well as long as all the other cleanup paths call directly into
__unschedule_txq(), that should remove stations from the scheduler
when
they disconnect etc.
Yes, the disconnect scenario is a bad example. My concern is, say, 
we
have 10 stations and only one of them is assigned a very small 
weight
compared with that of others. Suppose, after its chance of Tx, it is
most likely to be placed in the rightmost(still has some packets in
the
txq) and no more incoming data for it. The remaining packets in txq
will
be dropped due to timeout algo in codel(correct me if I am wrong) 
but
this empty txq will stay on the rbtree until other txqs get drained 
or
global vt catch up with its vt. The staying time could be long if
weight
is extremely small. Then do we need timer to check or any other 
better
solution?

Ah, I see what you mean. No, I don't think this will be a problem; 
the
scenario you're describing would play out like this:

1. Station ends transmitting, still has a single packet queued, gets
   moved to the end of the rbtree (and stays there for a while).

2. When we finally get to the point where this station gets another
   chance to transmit, the CoDel drop timer triggers and the last
packet
   is dropped[0]. This means that the queue will just be empty
   (and ieee80211_tx_dequeue() will return NULL).

3. Because the queue is empty, ieee80211_return_txq() will not put it
   back on the rbtree.

Crucially, in 2. the CoDel algorithm doesn't kick in until the point 
of
packet dequeue. But even if an empty queue stays on the rbtree for a
while, there is no harm in that: eventually it will get its turn, it
will turn out to be empty, and just be skipped over.
Then that will be fine. Thanks for the explanation of the dropping 
part
in CoDel algorithm.

Yup, think so. And you're welcome :)

The issue we need to be concerned about is the opposite: If we have a
queue that *does* have packets queued, but which is *not* scheduled 
for
transmission, that will stall TX.
Is it by design since its vt is more than global vt, right? The 
lattency
may somehow get impacted though.

Well, it should still stay on the rbtree as long as it has packets
queued. We don't have a check anywhere that reschedules TXQs whose v_t
drops below global v_t...

[0] CoDel in most cases only drops a single packet at a time, so it
will
not clear out an entire queue with multiple packets in one go. But 
you
are right that it could conceivably drop the last packet in a queue.

We only need to defer removal inside a single "scheduling round"
(i.e.,
between a pair of ieee80211_txq_schedule_start/end. So if we just
walk
the remove list in schedule_end() we should be enough, no?

Hmm, or maybe a simpler way to fix the original issue is just to 
have
unschedule_txq() update the schedule_pos() pointer?

I.e., unschedule_txq checks if the txq being removed is currently
being
pointed to by schedule_pos[ac], and if it is, it updates 
schedule_pos
to
be the rb_next of the current value?
Actually, if schedule_pos is updated to rb_next of the current 
value,
then in the next_txq() where we are going to use rb_next again and
finally pick the next node of the node we really want. Is it fine to
update schedule_pos to NULL?

Hmm, yeah, good point.

If we do end up setting schedule_pos to NULL in the middle of a
scheduling round, that will make next_txq() "start over", and do
another
loop through the whole thing. I guess we may be able hit a case where
things can oscillate back and forth between addition and removal
resulting in an infinite loop? Not sure, but at least I can't seem to
convince myself that this can't happen.

As the loop of next_txq under lock protection as below,

txq_schedule_start();
while(txq=next_txq()){
...
return_txq(txq);
}
txq_schedule_end();

I do not see any chance of addition, no?

As you noted in your other email, Felix reduced the locking. And yeah,
we need to rebase this series to also incorporate that. I figure I can
send an updated version of the first patch in the series once we've
worked out the remaining issues with your follow-up patches.

Oh, I was thinking we were discussing without locking reduced. Yes, I 
also agree there might be a case causing infinite loop. With locking 
reduced, the tree can be adjusted between next_txq() and return_txq() in 
the loop situation. For further discussion, let 's consider,
1) the tree starts like:
       A->B->C->D->E
2) then next_txq() returns A for dequeuing
3) driver dequeues A and draines A without any active txq locked meaning 
the tree could be changed upon Tx compeletion.
4) then in return_txq(), the tree could be,
       i   A->B->C->D->E （A is empty, and maybe soon be added back 
before the loop end）
       ii  B->C->A->D->E （A is empty, and maybe soon be added back 
before the loop end）
       iii B->C->D->E->A （A is empty, and maybe soon be added back 
before the loop end)

with this change:
 local->schedule_pos[ac] = rb_next(node) ?: rb_prev(node);

for case i, local->schedule_pos[ac] is rb_next(A) which is B, and in 
next_txq(), rb_next(B) is what we returns which actually is C and B is 
skipped, no?

Similiar for case ii, we skip B, C, D.

Also I am wondering if there will be some SMP issues relating with 
local->schedule_pos[ac].

In ath10k, we will usually push packets of first txq as many as we can
until it is drained and then move to the next one. So if a txq gets
removed in the return_txq, it should always be the leftmost. And
during this period, neither vt of any station or global vt can be
updated due to lock protection.


But in that case, we could fix it by just conditionally assigning
either
rb_next or rb_prev to the schedule_pos in unschedule_txq()? I.e.,
something like:

local->schedule_pos[ac] = rb_next(node) ?: rb_prev(node);
I am not sure I am getting your point. Still in next_txq,
schedule_pos[ac] will lead us to the next node of the one we want.

The logic in next_txq is different when schedule_pos[ac] is NULL, vs
when rb_next(schedule_pos[ac]) is NULL. The former restarts a new
scheduling round, while the latter ends the current round.

-Toke

--
Yibo






