If a uAPSD service period ends with an MMPDU, we currently just
send that MMPDU, but it obviously won't get the EOSP bit set as
it doesn't have a QoS header. This contradicts the standard, so
add a QoS-nulldata frame after the MMPDU to properly terminate
the service period with a frame that has EOSP set.

Also fix a bug wrt. the TID for the MMPDU, it shouldn't be set
to 0 unconditionally but use the actual TID that was assigned.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When a response for PS-Poll or a uAPSD trigger frame is sent, the
more-data bit should be set according to 802.11-2012 11.2.1.5 h),
meaning that it should indicate more data on the relevant ACs
(delivery-enabled or nondelivery-enabled for uAPSD or PS-Poll.)

In, for example, the following scenario:
 * 1 frame on VO queue (either in driver or in mac80211)
 * at least 1 frame on VI queue (in the driver)
 * both VO/VI are delivery-enabled
 * uAPSD trigger frame received

The more-data flag to the driver would not be set, even though
it should be.

While fixing this, I noticed that we should really release frames
from multiple ACs where there's data buffered in the driver for
the corresponding TIDs.

To address all this, restructure the code a bit to consider all
ACs if we only release driver frames or only buffered frames.
This also addresses the more-data bug described above as now the
TIDs will all be marked as released, so the driver will have to
check the number of frames.

While at it, clarify some code and comments and remove the found
variable, replacing it with the appropriate sw/hw release check.
Reported-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Using ffs() for the PS-Poll release TID is wrong, it will cause
frames to be released in order 0 1 2 3 4 5 6 7 instead of the
correct 7 6 5 4 3 0 2 1. Fix this by adding a new function that
implements "highest priority TID" properly.
Reported-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When the station's TX latency data structures need to be
allocated, handle failures properly and also free all the
structures if there are any other problems.

Move the allocation code up so that allocation failures
don't trigger rate control algorithm calls.
Reported-by: NZHAO Gang <gamerh2o@gmail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This is useful for debugging issues with drivers using this
function (erroneously), so add tracing for the API call.

Change-Id: Ice9d7eabb8fecbac188f0a741920d3488de700ec
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Teach sta_info_flush() to optionally also remove stations
from all VLANs associated with an AP interface to optimise
the station removal (in particular, synchronize_net().)

To not have to add the vlans argument throughout, do some
refactoring.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There's no reason to have one synchronize_net() for each
removed station, refactor the code slightly to have just
a single synchronize_net() for all stations.

Note that this is currently useless as hostapd removes
stations one by one and this coalescing never happens.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There's no reason to do this inside the sta key removal
since the keys can only be reached through the sta (and
not by the driver at all) so once the sta can no longer
be reached, the keys are safe.

This will allow further optimisation opportunities with
multiple stations.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

If we can assume that stations are never referenced by the
driver after sta_state returns (and this is true since the
previous iwlmvm patch and for all other drivers) then we
don't need to delay station destruction, and don't need to
play tricks with rcu_barrier() etc.

This should speed up some scenarios like hostapd shutdown.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The pointer should be cleared before synchronize_rcu() so that the
consequently dead station won't be found by any lookups in the TX
or RX paths.

Also check that the station is actually the one being removed, the
check is not needed because each 4-addr VLAN can only have a single
station and non-4-addr VLANs always have a NULL pointer there, but
the code is clearer this way (and we avoid the memory write.)
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Currently, mac80211 allows drivers to keep RCU-protected station
references that are cleared when the station is removed from the
driver and consequently needs to synchronize twice, once before
removing the station from the driver (so it can guarantee that
the station is no longer used in TX towards the driver) and once
after the station is removed from the driver.

Add a new pre-RCU-synchronisation station removal operation to
the API to allow drivers to clear/invalidate their RCU-protected
station pointers before the RCU synchronisation.

This will allow removing the second synchronisation by changing
the driver API so that the driver may no longer assume a valid
RCU-protected pointer after sta_remove/sta_state returns.

The alternative to this would be to synchronize_rcu() in all the
drivers that currently rely on this behaviour (only iwlmvm) but
that would defeat the purpose.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This shouldn't be happening, so warn if it does, since we'd
then leak a station entry.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Measure TX latency and jitter statistics per station per TID.
These Measurements are disabled by default and can be enabled
via debugfs.

Features included for each station's TID:

1. Keep count of the maximum and average latency of Tx frames.
2. Keep track of many frames arrived in a specific time range
   (need to enable through debugfs and configure the bins ranges)
Signed-off-by: NMatti Gottlieb <matti.gottlieb@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Add a new field to ieee80211_chanctx_conf to indicate
the min required channel configuration.

Tuning to a narrower channel might help reducing
the noise level and saving some power.

The min required channel definition is the max of
all min required channel definitions of the interfaces
bound to this channel context.

In AP mode, use 20MHz when there are no connected station.
When a new station is added/removed, calculate the new max
bandwidth supported by any of the stations (e.g. 80MHz when
80MHz and 40MHz stations are connected).

In other cases, simply use bss_conf.chandef as the
min required chandef.

Notify drivers about changes to this field by calling
drv_change_chanctx with a new CHANGE_MIN_WIDTH notification.
Signed-off-by: NEliad Peller <eliad@wizery.com>
Reviewed-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Use put_unaligned_le16 in mesh_plink_frame_tx.
Signed-off-by: NChun-Yeow Yeoh <yeohchunyeow@gmail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

802.11-2012 13.3.1 implicitly limits the mesh local link
ID range to that of AID, since for mesh PS the local link
ID must be indicated in the TIM IE, which only holds
IEEE80211_MAX_AID bits.

Also the code was allowing a local link ID of 0, but this
is not correct since that TIM bit is used for indicating
buffered mcast frames.

Generate a random, unique, link ID from 1 - 2007, and drop
a modulo conversion for the local link ID, but keep it for
the peer link ID in case he chose something > MAX_AID.
Signed-off-by: NThomas Pedersen <thomas@cozybit.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When the driver requests to move to STATIC or DYNAMIC SMPS,
we send an action frame to each associated station and
reconfigure the channel context / driver.
Of course, non-MIMO stations are ignored.

The beacon isn't updated. The association response will
include the original capabilities. Stations that associate
while in non-OFF SMPS mode will get an action frame right
after association to inform them about our current state.
Note that we wait until the end of the EAPOL. Sending an
action frame before the EAPOL is finished can be an issue
for a few clients. Clients aren't likely to send EAPOL
frames in MIMO anyway.

When the SMPS configuration gets more permissive (e.g.
STATIC -> OFF), we don't wake up stations that are asleep
We remember that they don't know about the change and send
the action frame when they wake up.

When the SMPS configuration gets more restrictive (e.g.
OFF -> STATIC), we set the TIM bit for every sleeping STA.
uAPSD stations might send MIMO until they poll the action
frame, but this is for a short period of time.
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
[fix vht streams loop, initialisation]
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Previously the default mesh STA nonpeer power mode was
UNKNOWN (0) make the default mesh STA power mode ACTIVE,
to prevent unnecessary  frame buffering while peering is
not yet complete. Fixes a panic in ath9k_htc when adding
stations from userspace, and mcast buffered frames are
later released.

Thanks to Bob Copeland for his help debugging this.
Signed-off-by: NThomas Pedersen <thomas@cozybit.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Ben reports that kmemleak is saying TX aggregation TID
structs are leaked. Given his workload, I suspect that
they're leaked because stations are destroyed before
their aggregation sessions get a chance to start. Fix
this by simply freeing structs that are not used yet.
Reported-by: NBen Greear <greearb@candelatech.com>
Tested-by: NBen Greear <greearb@candelatech.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Signed-off-by: NFelix Fietkau <nbd@openwrt.org>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Signed-off-by: NFelix Fietkau <nbd@openwrt.org>
[fix unit documentation]
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Instead of calling synchronize_net() for every key
on an interface or when a station is removed, do it
only once for all keys in both of these cases.

As a side-effect, removing station keys now always
calls synchronize_net() even if there are no keys,
which fixes an issue with station removal happening
in the driver while the station could still be used
for TX.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Some key function don't exist without underscores, so
remove the underscores from those.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

It is possible that sta_info_recalc_tim() is called consecutively
without changing the station's tim bit. In such cases there is no
need to call the driver's set_tim() callback.
Signed-off-by: NIlan Peer <ilan.peer@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The irqsafe version ieee80211_sta_eosp_irqsafe() exists, but
drivers must not mix calls to any irqsafe/non-irqsafe function.
Both ath9k and iwlwifi, the likely first users of this interface,
use non-irqsafe RX/TX/TX status so must also use a non-irqsafe
version of this function. Since no driver uses the _irqsafe()
version, remove that.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

If there are keys left during station removal, then a
synchronize_net() will be done (for each key, I have a
patch to address this for 3.10), otherwise it won't be
done at all which causes issues because the station
could be used for TX while it's being removed from the
driver -- that might confuse the driver.

Fix this by always doing synchronize_net() if no key
was present any more.

Cc: stable@vger.kernel.org
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There is a corner case which wasn't being covered:
userspace may authenticate and allocate stations,
but still leave the peering up to the kernel.

Initialize the peering timer if the MPM is not in
userspace, in a path which is taken by both the kernel and
userspace when allocating stations.
Signed-off-by: NThomas Pedersen <thomas@cozybit.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

During roaming, the crypto_tx_tailroom_needed_cnt counter
will often take values 2,1,0,1,2 because first keys are
removed and then new keys are added. This is inefficient
because during the 0->1 transition, synchronize_net must
be called to avoid packet races, although typically no
packets would be flowing during that time.

To avoid that, defer the decrement (2->1, 1->0) when keys
are removed (by half a second). This means the counter
will really have the values 2,2,2,3,4 ... 2, thus never
reaching 0 and having to do the 0->1 transition.

Note that this patch entirely disregards the drivers for
which this optimisation was done to start with, for them
the key removal itself will be expensive because it has
to synchronize_net() after the counter is incremented to
remove the key from HW crypto. For them the sequence will
look like this: 0,1,0,1,0,1,0,1,0 (*) which is clearly a
lot more inefficient. This could be addressed separately,
during key removal the 0->1->0 sequence isn't necessary.

(*) it starts at 0 because HW crypto is on, then goes to
    1 when HW crypto is disabled for a key, then back to
    0 because the key is deleted; this happens for both
    keys in the example. When new keys are added, it goes
    to 1 first because they're added in software; when a
    key is moved to hardware it goes back to 0
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Instead of modifying the HT SMPS capability field
for stations, track the SMPS mode explicitly in a
new field in the station struct and use it in the
drivers that care about it. This simplifies the
code using it.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There's no need to use _irqsave() as the lock
is never used in interrupt context.

This also fixes a problem in the iwlwifi MVM
driver that calls spin_unlock_bh() within its
set_tim() callback.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

In my commit 1672c0e3
("mac80211: start auth/assoc timeout on frame status")
I broke auth/assoc timeout handling: in case we wait
for the TX status, it now leaves the timeout field set
to 0, which is a valid time and can compare as being
before now ("jiffies"). Thus, if the work struct runs
for some other reason, the auth/assoc is treated as
having timed out.

Fix this by introducing a separate "timeout_started"
variable that tracks whether the timeout has started
and is checked before timing out.

Additionally, for proper TX status handling the change
requires that the skb->dev pointer is set up for all
the frames, so set it up for all frames in mac80211.
Reported-by: NWojciech Dubowik <Wojciech.Dubowik@neratec.com>
Tested-by: NWojciech Dubowik <Wojciech.Dubowik@neratec.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The patch "mac80211: clean up mesh sta allocation warning"
moved some mesh initialization into a path which is only
called when the kernel handles peering. This causes a hang
when mac80211 tries to clean up a userspace-allocated
station entry and delete a timer which has never been
initialized.

To avoid this, only do any mesh sta peering teardown if
the kernel is actually handling it.

The same is true when quiescing before suspend.
Signed-off-by: NThomas Pedersen <thomas@cozybit.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Add routines to
- maintain a PS mode for each peer and a non-peer PS mode
- indicate own PS mode in transmitted frames
- track neighbor STAs power modes
- buffer frames when neighbors are in PS mode
- add TIM and Awake Window IE to beacons
- release frames in Mesh Peer Service Periods

Add local_pm to sta_info to represent the link-specific power
mode at this station towards the remote station. When a peer
link is established, use the default power mode stored in mesh
config. Update the PS status if the peering status of a neighbor
changes.
Maintain a mesh power mode for non-peer mesh STAs. Set the
non-peer power mode to active mode during peering. Authenticated
mesh peering is currently not working when either node is
configured to be in power save mode.

Indicate the current power mode in transmitted frames. Use QoS
Nulls to indicate mesh power mode transitions.
For performance reasons, calls to the function setting the frame
flags are placed in HWMP routing routines, as there the STA
pointer is already available.

Add peer_pm to sta_info to represent the peer's link-specific
power mode towards the local station. Add nonpeer_pm to
represent the peer's power mode towards all non-peer stations.
Track power modes based on received frames.

Add the ps_data structure to ieee80211_if_mesh (for TIM map, PS
neighbor counter and group-addressed frame buffer).

Set WLAN_STA_PS flag for STA in PS mode to use the unicast frame
buffering routines in the tx path. Update num_sta_ps to buffer
and release group-addressed frames after DTIM beacons.

Announce the awake window duration in beacons if in light or
deep sleep mode towards any peer or non-peer. Create a TIM IE
similarly to AP mode and add it to mesh beacons. Parse received
Awake Window IEs and check TIM IEs for buffered frames.

Release frames towards peers in mesh Peer Service Periods. Use
the corresponding trigger frames and monitor the MPSP status.
Append a QoS Null as trigger frame if neccessary to properly end
the MPSP. Currently, in HT channels MPSPs behave imperfectly and
show large delay spikes and frame losses.
Signed-off-by: NMarco Porsch <marco@cozybit.com>
Signed-off-by: NIvan Bezyazychnyy <ivan.bezyazychnyy@gmail.com>
Signed-off-by: NMike Krinkin <krinkin.m.u@gmail.com>
Signed-off-by: NMax Filippov <jcmvbkbc@gmail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This refactoring fixes a "scheduling while atomic" warning
when allocating a mesh station entry while holding the RCU
read lock. Fix this by creating a new function
mesh_sta_info_get(), which correctly handles the locking
and returns under RCU.

Also move some unnecessarily #ifdefed mesh station init
code from sta_info_alloc() to __mesh_sta_info_alloc().
Signed-off-by: NThomas Pedersen <thomas@cozybit.com>
[change code flow to make sparse happy]
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The initiator/tx doesn't really identify why an
aggregation session is stopped, give a reason
for stopping that more clearly identifies what's
going on. This will help tell the driver clearly
what is expected of it.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

If there are VLANs, stopping an AP is inefficient as it
calls rcu_barrier() once for each interface (the VLANs
and the AP itself). Optimise this by moving rcu_barrier()
out of the station cleanups and calling it only once for
all interfaces combined.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The last fixes re-added the RCU synchronize penalty
on roaming to fix the races. Split up sta_info_flush()
now to get rid of that again, and let managed mode
(and only it) delay the actual destruction.
Tested-by: NBen Greear <greearb@candelatech.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When all interfaces have been removed, there can't
be any stations left over, so there's no need to
flush again. Remove this, and all code associated
with it, which also simplifies the function.
Tested-by: NBen Greear <greearb@candelatech.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This is a very old bug, but there's nothing that prevents the
timer from running while the module is being removed when we
only do del_timer() instead of del_timer_sync().

The timer should normally not be running at this point, but
it's not clearly impossible (or we could just remove this.)

Cc: stable@vger.kernel.org
Tested-by: NBen Greear <greearb@candelatech.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Unfortunately, commit b22cfcfc, intended to speed up roaming
by avoiding the synchronize_rcu() broke AP/mesh modes as it moved
some code into that work item that will still call into the driver
at a time where it's no longer expected to handle this: after the
AP or mesh has been stopped.

To fix this problem remove the per-station work struct, maintain a
station cleanup list instead and flush this list when stations are
flushed. To keep this patch smaller for stable, do this when the
stations are flushed (sta_info_flush()). This unfortunately brings
back the original roaming delay; I'll fix that again in a separate
patch.

Also, Ben reported that the original commit could sometimes (with
many interfaces) cause long delays when an interface is set down,
due to blocking on flush_workqueue(). Since we now maintain the
cleanup list, this particular change of the original patch can be
reverted.

Cc: stable@vger.kernel.org [3.7]
Reported-by: NBen Greear <greearb@candelatech.com>
Tested-by: NBen Greear <greearb@candelatech.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>