| Message ID | 1313053697-19544-1-git-send-email-helmut.schaa@googlemail.com (mailing list archive) |
|---|---|
| State | Not Applicable, archived |
| Headers | show |
FWIW, this is exactly the behaviour I'm currently writing into FreeBSD's net80211/ath ADDBA handling on BAR TX failure. (When aggregation is enabled, whether or not it's an A-MPDU.) So I think it's a good idea, as the only way subsequent packets are going to flow is if the sender has some subsequent frames to send that cause the BAW to slide along. If it's (for example) some interactive TCP or UDP, I can certainly see the session hanging. I'm not sending BAR's yet on ADDBA session TX failures (it's actually what I'm just about to work on), and I came to the same conclusion as you when I noticed ICMP pings stopping when a frame wasn't successfully received. So FWIW, +1 from me. Adrian On 11 August 2011 17:08, Helmut Schaa <helmut.schaa@googlemail.com> wrote: > As described at [1] some STAs (i.e. Intel 5100 on Windows) can end up > correctly BlockAcking incoming frames without delivering them to user > space if a AMPDU subframe got lost and its reorder buffer isn't flushed > by a BlockAckReq. This in turn results in up to 64 frames being stuck > in the reorder buffer. > > Accroding to 802.11n-2009 it is not necessary to send a BAR to flush > the receipients RX reorder buffer but we still do that to be polite. > > However, assume the following frame exchange: > > AP -> STA, AMPDU (failed) > AP -> STA, BAR (failed) > > The client in question then ends up in the same situation and won't > deliver frames to userspace anymore since we weren't able to flush > its reorder buffer. > > This is not a hypothetical situation but I was able to observe this > exact behavior during a stress test between a rt2800pci AP and a Intel > 5100 Windows client. > > In order to work around this issue just tear down the BA session as > soon as a BAR failed to be TX'ed. > > [1] http://comments.gmane.org/gmane.linux.kernel.wireless.general/66867 > > Signed-off-by: Helmut Schaa <helmut.schaa@googlemail.com> > --- > > IMHO the Windows driver is just buggy and should be fixed to use a > reasonable timeout for flushing its reorder buffer but the described > behavior doesn't appear with the Ralink Legacy drivers for example since > they trigger a tear down of the BA session in several other situations > as well (a single failed AMPDU :) for example) and thus don't end up in > this situation. > > Johannes, feel free to NACK this patch as it really is just a > workaround for buggy clients but I'd say it still makes sense to fall > back to non-aggregated frames in such a situation. Furthermore, this > situation is unlikely to happen very often but as written before I was > able to reproduce it a couple of times. > > Thanks, > Helmut > > net/mac80211/status.c | 13 +++++++++++++ > 1 files changed, 13 insertions(+), 0 deletions(-) > > diff --git a/net/mac80211/status.c b/net/mac80211/status.c > index 1658efa..6c4b728 100644 > --- a/net/mac80211/status.c > +++ b/net/mac80211/status.c > @@ -187,6 +187,7 @@ void ieee80211_tx_status(struct ieee80211_hw *hw, struct sk_buff *skb) > int rates_idx = -1; > bool send_to_cooked; > bool acked; > + struct ieee80211_bar *bar; > > for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) { > if (info->status.rates[i].idx < 0) { > @@ -243,6 +244,18 @@ void ieee80211_tx_status(struct ieee80211_hw *hw, struct sk_buff *skb) > tid, ssn); > } > > + if (!acked && ieee80211_is_back_req(fc)) { > + /* > + * BAR failed, let's tear down the BA session as a > + * last resort as some STAs (Intel 5100 on Windows) > + * can get stuck when the BA window isn't flushed > + * correctly. > + */ > + bar = (struct ieee80211_bar *) skb->data; > + ieee80211_stop_tx_ba_session(&sta->sta, > + bar->control >> 12 & 0xf); > + } > + > if (info->flags & IEEE80211_TX_STAT_TX_FILTERED) { > ieee80211_handle_filtered_frame(local, sta, skb); > rcu_read_unlock(); > -- > 1.7.3.4 > > -- > To unsubscribe from this list: send the line "unsubscribe linux-wireless" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- To unsubscribe from this list: send the line "unsubscribe linux-wireless" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, 2011-08-11 at 11:08 +0200, Helmut Schaa wrote: > As described at [1] some STAs (i.e. Intel 5100 on Windows) can end up > correctly BlockAcking incoming frames without delivering them to user > space if a AMPDU subframe got lost and its reorder buffer isn't flushed > by a BlockAckReq. This in turn results in up to 64 frames being stuck > in the reorder buffer. > > Accroding typo > to 802.11n-2009 it is not necessary to send a BAR to flush > the receipients RX reorder buffer but we still do that to be polite. typo > IMHO the Windows driver is just buggy and should be fixed to use a > reasonable timeout for flushing its reorder buffer but the described > behavior doesn't appear with the Ralink Legacy drivers for example since > they trigger a tear down of the BA session in several other situations > as well (a single failed AMPDU :) for example) and thus don't end up in > this situation. > > Johannes, feel free to NACK this patch as it really is just a > workaround for buggy clients but I'd say it still makes sense to fall > back to non-aggregated frames in such a situation. Furthermore, this > situation is unlikely to happen very often but as written before I was > able to reproduce it a couple of times. Seems ok to me, hopefully won't happen often :) > + if (!acked && ieee80211_is_back_req(fc)) { > + /* > + * BAR failed, let's tear down the BA session as a > + * last resort as some STAs (Intel 5100 on Windows) > + * can get stuck when the BA window isn't flushed > + * correctly. > + */ > + bar = (struct ieee80211_bar *) skb->data; > + ieee80211_stop_tx_ba_session(&sta->sta, > + bar->control >> 12 & 0xf); > + } Hmm, that shift & mask makes me think twice, are there constants, and maybe there should be some parentheses? johannes -- To unsubscribe from this list: send the line "unsubscribe linux-wireless" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Aug 11, 2011 at 2:58 PM, Johannes Berg <johannes@sipsolutions.net> wrote: >> + if (!acked && ieee80211_is_back_req(fc)) { >> + /* >> + * BAR failed, let's tear down the BA session as a >> + * last resort as some STAs (Intel 5100 on Windows) >> + * can get stuck when the BA window isn't flushed >> + * correctly. >> + */ >> + bar = (struct ieee80211_bar *) skb->data; >> + ieee80211_stop_tx_ba_session(&sta->sta, >> + bar->control >> 12 & 0xf); >> + } > > Hmm, that shift & mask makes me think twice, are there constants, and > maybe there should be some parentheses? This just masks out the TID associated to this BA agreement and the shift has a higher precedence then the bitwise &. We don't have a suitable constant yet, a hardcoded 12 is also used in ieee80211_send_bar. Hence, I guess a define would be suitable here. I'll resend with the fixed typos and replace the 12 with a define. Thanks, Helmut -- To unsubscribe from this list: send the line "unsubscribe linux-wireless" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/net/mac80211/status.c b/net/mac80211/status.c index 1658efa..6c4b728 100644 --- a/net/mac80211/status.c +++ b/net/mac80211/status.c @@ -187,6 +187,7 @@ void ieee80211_tx_status(struct ieee80211_hw *hw, struct sk_buff *skb) int rates_idx = -1; bool send_to_cooked; bool acked; + struct ieee80211_bar *bar; for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) { if (info->status.rates[i].idx < 0) { @@ -243,6 +244,18 @@ void ieee80211_tx_status(struct ieee80211_hw *hw, struct sk_buff *skb) tid, ssn); } + if (!acked && ieee80211_is_back_req(fc)) { + /* + * BAR failed, let's tear down the BA session as a + * last resort as some STAs (Intel 5100 on Windows) + * can get stuck when the BA window isn't flushed + * correctly. + */ + bar = (struct ieee80211_bar *) skb->data; + ieee80211_stop_tx_ba_session(&sta->sta, + bar->control >> 12 & 0xf); + } + if (info->flags & IEEE80211_TX_STAT_TX_FILTERED) { ieee80211_handle_filtered_frame(local, sta, skb); rcu_read_unlock();
As described at [1] some STAs (i.e. Intel 5100 on Windows) can end up correctly BlockAcking incoming frames without delivering them to user space if a AMPDU subframe got lost and its reorder buffer isn't flushed by a BlockAckReq. This in turn results in up to 64 frames being stuck in the reorder buffer. Accroding to 802.11n-2009 it is not necessary to send a BAR to flush the receipients RX reorder buffer but we still do that to be polite. However, assume the following frame exchange: AP -> STA, AMPDU (failed) AP -> STA, BAR (failed) The client in question then ends up in the same situation and won't deliver frames to userspace anymore since we weren't able to flush its reorder buffer. This is not a hypothetical situation but I was able to observe this exact behavior during a stress test between a rt2800pci AP and a Intel 5100 Windows client. In order to work around this issue just tear down the BA session as soon as a BAR failed to be TX'ed. [1] http://comments.gmane.org/gmane.linux.kernel.wireless.general/66867 Signed-off-by: Helmut Schaa <helmut.schaa@googlemail.com> --- IMHO the Windows driver is just buggy and should be fixed to use a reasonable timeout for flushing its reorder buffer but the described behavior doesn't appear with the Ralink Legacy drivers for example since they trigger a tear down of the BA session in several other situations as well (a single failed AMPDU :) for example) and thus don't end up in this situation. Johannes, feel free to NACK this patch as it really is just a workaround for buggy clients but I'd say it still makes sense to fall back to non-aggregated frames in such a situation. Furthermore, this situation is unlikely to happen very often but as written before I was able to reproduce it a couple of times. Thanks, Helmut net/mac80211/status.c | 13 +++++++++++++ 1 files changed, 13 insertions(+), 0 deletions(-)