Message ID | 20190824134857.4094-1-erik.stromdahl@gmail.com (mailing list archive) |
---|---|
State | Changes Requested |
Delegated to: | Kalle Valo |
Headers | show |
Series | ath10k: remove TX lock from ath10k_htt_tx_inc_pending | expand |
> -----Original Message----- > From: ath10k <ath10k-bounces@lists.infradead.org> On Behalf Of Erik > Stromdahl > Sent: Saturday, August 24, 2019 9:49 PM > To: Kalle Valo <kvalo@qca.qualcomm.com>; linux-wireless@vger.kernel.org; > ath10k@lists.infradead.org > Cc: Erik Stromdahl <erik.stromdahl@gmail.com> > Subject: [EXT] [PATCH] ath10k: remove TX lock from > ath10k_htt_tx_inc_pending > > This commit removes the call to ath10k_mac_tx_lock() from > ath10k_htt_tx_inc_pending() in case the high water mark is reached. > > ath10k_mac_tx_lock() calls ieee80211_stop_queues() in order to stop > mac80211 from pushing more TX data to the driver (this is the TX lock). > > If a driver is trying to fetch an skb from a queue while the queue is > stopped, ieee80211_tx_dequeue() will return NULL. > > So, in ath10k_mac_tx_push_txq(), there is a risk that the call to > ath10k_htt_tx_inc_pending() results in a stop of the mac80211 TX queues > just before the skb is fetched. > > This will cause ieee80211_tx_dequeue() to return NULL and > ath10k_mac_tx_push_txq() to exit prematurely and return -ENOENT. > Before the function returns ath10k_htt_tx_dec_pending() will be called. > This call will re-enable the TX queues through ath10k_mac_tx_unlock(). > When ath10k_mac_tx_push_txq() has returned, the TX queue will be > returned back to mac80211 with ieee80211_return_txq() without the skb > being properly consumed. > > Since the TX queues were re-enabled in the error exit path of > ath10k_mac_tx_push_txq(), mac80211 can continue pushing data to the > driver. If the hardware does not consume the data, the above mentioned > case will be repeated over and over. > > A case when the hardware is not able to transmit the data from the host > is when a STA has been dis-associated from an AP and has not yet been > able to re-associate. In this case there will be no TX_COMPL_INDs from > the hardware, resulting in the TX counter not be decremented. > > This phenomenon has been observed in both a real and a test setup. > > In order to fix this, the actual TX locking (the call to > ath10k_mac_tx_lock()) was removed from ath10k_htt_tx_inc_pending(). > Instead, ath10k_mac_tx_lock() is called separately after the skb has > been fetched (after the call to ieee80211_tx_dequeue()). At this point > it is OK to stop the queues. Is this patch will impact throughput? > > > _______________________________________________ > ath10k mailing list > ath10k@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/ath10k
On 8/26/19 4:44 AM, Wen Gong wrote: >> -----Original Message----- >> From: ath10k <ath10k-bounces@lists.infradead.org> On Behalf Of Erik >> Stromdahl >> Sent: Saturday, August 24, 2019 9:49 PM >> To: Kalle Valo <kvalo@qca.qualcomm.com>; linux-wireless@vger.kernel.org; >> ath10k@lists.infradead.org >> Cc: Erik Stromdahl <erik.stromdahl@gmail.com> >> Subject: [EXT] [PATCH] ath10k: remove TX lock from >> ath10k_htt_tx_inc_pending >> >> This commit removes the call to ath10k_mac_tx_lock() from >> ath10k_htt_tx_inc_pending() in case the high water mark is reached. >> >> ath10k_mac_tx_lock() calls ieee80211_stop_queues() in order to stop >> mac80211 from pushing more TX data to the driver (this is the TX lock). >> >> If a driver is trying to fetch an skb from a queue while the queue is >> stopped, ieee80211_tx_dequeue() will return NULL. >> >> So, in ath10k_mac_tx_push_txq(), there is a risk that the call to >> ath10k_htt_tx_inc_pending() results in a stop of the mac80211 TX queues >> just before the skb is fetched. >> >> This will cause ieee80211_tx_dequeue() to return NULL and >> ath10k_mac_tx_push_txq() to exit prematurely and return -ENOENT. >> Before the function returns ath10k_htt_tx_dec_pending() will be called. >> This call will re-enable the TX queues through ath10k_mac_tx_unlock(). >> When ath10k_mac_tx_push_txq() has returned, the TX queue will be >> returned back to mac80211 with ieee80211_return_txq() without the skb >> being properly consumed. >> >> Since the TX queues were re-enabled in the error exit path of >> ath10k_mac_tx_push_txq(), mac80211 can continue pushing data to the >> driver. If the hardware does not consume the data, the above mentioned >> case will be repeated over and over. >> >> A case when the hardware is not able to transmit the data from the host >> is when a STA has been dis-associated from an AP and has not yet been >> able to re-associate. In this case there will be no TX_COMPL_INDs from >> the hardware, resulting in the TX counter not be decremented. >> >> This phenomenon has been observed in both a real and a test setup. >> >> In order to fix this, the actual TX locking (the call to >> ath10k_mac_tx_lock()) was removed from ath10k_htt_tx_inc_pending(). >> Instead, ath10k_mac_tx_lock() is called separately after the skb has >> been fetched (after the call to ieee80211_tx_dequeue()). At this point >> it is OK to stop the queues. > Is this patch will impact throughput? No, I haven't seen any performance degradation. I can't see that it should impact the throughput even in theory (rather the opposite actually since we reduce the likelihood of dropping packets).
Erik Stromdahl <erik.stromdahl@gmail.com> wrote: > This commit removes the call to ath10k_mac_tx_lock() from > ath10k_htt_tx_inc_pending() in case the high water mark is reached. > > ath10k_mac_tx_lock() calls ieee80211_stop_queues() in order to stop > mac80211 from pushing more TX data to the driver (this is the TX lock). > > If a driver is trying to fetch an skb from a queue while the queue is > stopped, ieee80211_tx_dequeue() will return NULL. > > So, in ath10k_mac_tx_push_txq(), there is a risk that the call to > ath10k_htt_tx_inc_pending() results in a stop of the mac80211 TX queues > just before the skb is fetched. > > This will cause ieee80211_tx_dequeue() to return NULL and > ath10k_mac_tx_push_txq() to exit prematurely and return -ENOENT. > Before the function returns ath10k_htt_tx_dec_pending() will be called. > This call will re-enable the TX queues through ath10k_mac_tx_unlock(). > When ath10k_mac_tx_push_txq() has returned, the TX queue will be > returned back to mac80211 with ieee80211_return_txq() without the skb > being properly consumed. > > Since the TX queues were re-enabled in the error exit path of > ath10k_mac_tx_push_txq(), mac80211 can continue pushing data to the > driver. If the hardware does not consume the data, the above mentioned > case will be repeated over and over. > > A case when the hardware is not able to transmit the data from the host > is when a STA has been dis-associated from an AP and has not yet been > able to re-associate. In this case there will be no TX_COMPL_INDs from > the hardware, resulting in the TX counter not be decremented. > > This phenomenon has been observed in both a real and a test setup. > > In order to fix this, the actual TX locking (the call to > ath10k_mac_tx_lock()) was removed from ath10k_htt_tx_inc_pending(). > Instead, ath10k_mac_tx_lock() is called separately after the skb has > been fetched (after the call to ieee80211_tx_dequeue()). At this point > it is OK to stop the queues. > > Signed-off-by: Erik Stromdahl <erik.stromdahl@gmail.com> What hardware and firmware versions did you test this? Please always add that to the commit log. As Erik mostly works on SDIO I assume PCI is not that well tested. Has anyone else tried this?
diff --git a/drivers/net/wireless/ath/ath10k/htt_tx.c b/drivers/net/wireless/ath/ath10k/htt_tx.c index 2ef717f18795..32fd71f28ef9 100644 --- a/drivers/net/wireless/ath/ath10k/htt_tx.c +++ b/drivers/net/wireless/ath/ath10k/htt_tx.c @@ -157,8 +157,6 @@ int ath10k_htt_tx_inc_pending(struct ath10k_htt *htt) return -EBUSY; htt->num_pending_tx++; - if (htt->num_pending_tx == htt->max_num_pending_tx) - ath10k_mac_tx_lock(htt->ar, ATH10K_TX_PAUSE_Q_FULL); return 0; } diff --git a/drivers/net/wireless/ath/ath10k/mac.c b/drivers/net/wireless/ath/ath10k/mac.c index c1bf1167a2ce..8d357c79ca59 100644 --- a/drivers/net/wireless/ath/ath10k/mac.c +++ b/drivers/net/wireless/ath/ath10k/mac.c @@ -4078,6 +4078,11 @@ int ath10k_mac_tx_push_txq(struct ieee80211_hw *hw, return -ENOENT; } + spin_lock_bh(&ar->htt.tx_lock); + if (htt->num_pending_tx == htt->max_num_pending_tx) + ath10k_mac_tx_lock(ar, ATH10K_TX_PAUSE_Q_FULL); + spin_unlock_bh(&ar->htt.tx_lock); + airtime = ath10k_mac_update_airtime(ar, txq, skb); ath10k_mac_tx_h_fill_cb(ar, vif, txq, skb, airtime); @@ -4370,6 +4375,9 @@ static void ath10k_mac_op_tx(struct ieee80211_hw *hw, return; } + if (htt->num_pending_tx == htt->max_num_pending_tx) + ath10k_mac_tx_lock(ar, ATH10K_TX_PAUSE_Q_FULL); + ret = ath10k_htt_tx_mgmt_inc_pending(htt, is_mgmt, is_presp); if (ret) { ath10k_dbg(ar, ATH10K_DBG_MAC, "failed to increase tx mgmt pending count: %d, dropping\n",
This commit removes the call to ath10k_mac_tx_lock() from ath10k_htt_tx_inc_pending() in case the high water mark is reached. ath10k_mac_tx_lock() calls ieee80211_stop_queues() in order to stop mac80211 from pushing more TX data to the driver (this is the TX lock). If a driver is trying to fetch an skb from a queue while the queue is stopped, ieee80211_tx_dequeue() will return NULL. So, in ath10k_mac_tx_push_txq(), there is a risk that the call to ath10k_htt_tx_inc_pending() results in a stop of the mac80211 TX queues just before the skb is fetched. This will cause ieee80211_tx_dequeue() to return NULL and ath10k_mac_tx_push_txq() to exit prematurely and return -ENOENT. Before the function returns ath10k_htt_tx_dec_pending() will be called. This call will re-enable the TX queues through ath10k_mac_tx_unlock(). When ath10k_mac_tx_push_txq() has returned, the TX queue will be returned back to mac80211 with ieee80211_return_txq() without the skb being properly consumed. Since the TX queues were re-enabled in the error exit path of ath10k_mac_tx_push_txq(), mac80211 can continue pushing data to the driver. If the hardware does not consume the data, the above mentioned case will be repeated over and over. A case when the hardware is not able to transmit the data from the host is when a STA has been dis-associated from an AP and has not yet been able to re-associate. In this case there will be no TX_COMPL_INDs from the hardware, resulting in the TX counter not be decremented. This phenomenon has been observed in both a real and a test setup. In order to fix this, the actual TX locking (the call to ath10k_mac_tx_lock()) was removed from ath10k_htt_tx_inc_pending(). Instead, ath10k_mac_tx_lock() is called separately after the skb has been fetched (after the call to ieee80211_tx_dequeue()). At this point it is OK to stop the queues. Signed-off-by: Erik Stromdahl <erik.stromdahl@gmail.com> --- drivers/net/wireless/ath/ath10k/htt_tx.c | 2 -- drivers/net/wireless/ath/ath10k/mac.c | 8 ++++++++ 2 files changed, 8 insertions(+), 2 deletions(-)