Mesh HWMP module will be able to rely on the HW
RC algorithm if it exists, for path metric calculations.

This allows the metric calculation mechanism to calculate
a correct metric, based on PER and last TX rate both via
HW RC algorithm if it exists or via parameters collected
by the SW.
Signed-off-by: NMaxim Altshul <maxim.altshul@ti.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

mac80211's software queues were designed to work
very closely with device tx queues. They are
required to make use of 802.11 packet aggregation
easily and efficiently.

Due to the way 802.11 aggregation is designed it
only makes sense to keep fair queuing as close to
hardware as possible to reduce induced latency and
inertia and provide the best flow responsiveness.

This change doesn't translate directly to
immediate and significant gains. End result
depends on driver's induced latency. Best results
can be achieved if driver keeps its own tx
queue/fifo fill level to a minimum.
Signed-off-by: NMichal Kazior <michal.kazior@tieto.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Qdiscs are designed with no regard to 802.11
aggregation requirements and hand out
packet-by-packet with no guarantee they are
destined to the same tid. This does more bad than
good no matter how fairly a given qdisc may behave
on an ethernet interface.

Software queuing used per-AC netdev subqueue
congestion control whenever a global AC limit was
hit. This meant in practice a single station or
tid queue could starve others rather easily. This
could resonate with qdiscs in a bad way or could
just end up with poor aggregation performance.
Increasing the AC limit would increase induced
latency which is also bad.

Disabling qdiscs by default and performing
taildrop instead of netdev subqueue congestion
control on the other hand makes it possible for
tid queues to fill up "in the meantime" while
preventing stations starving each other.

This increases aggregation opportunities and
should allow software queuing based drivers
achieve better performance by utilizing airtime
more efficiently with big aggregates.
Signed-off-by: NMichal Kazior <michal.kazior@tieto.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Requires software tx queueing and fast-xmit support. For good
performance, drivers need frag_list support as well. This avoids the
need for copying data of aggregated frames. Running without it is only
supported for debugging purposes.

To avoid performance and packet size issues, the rate control module or
driver needs to limit the maximum A-MSDU size by setting
max_rc_amsdu_len in struct ieee80211_sta.
Signed-off-by: NFelix Fietkau <nbd@openwrt.org>
[fix locking issue]
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

If the driver advertises the new HW flag USE_RSS, make the
station statistics on the fast-rx path per-CPU. This will
enable calling the RX in parallel, only hitting locking or
shared cachelines when the fast-RX path isn't available.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The regular RX path has a lot of code, but with a few
assumptions on the hardware it's possible to reduce the
amount of code significantly. Currently the assumptions
on the driver are the following:
 * hardware/driver reordering buffer (if supporting aggregation)
 * hardware/driver decryption & PN checking (if using encryption)
 * hardware/driver did de-duplication
 * hardware/driver did A-MSDU deaggregation
 * AP_LINK_PS is used (in AP mode)
 * no client powersave handling in mac80211 (in client mode)

of which some are actually checked per packet:
 * de-duplication
 * PN checking
 * decryption
and additionally packets must
 * not be A-MSDU (have been deaggregated by driver/device)
 * be data packets
 * not be fragmented
 * be unicast
 * have RFC 1042 header

Additionally dynamically we assume:
 * no encryption or CCMP/GCMP, TKIP/WEP/other not allowed
 * station must be authorized
 * 4-addr format not enabled

Some data needed for the RX path is cached in a new per-station
"fast_rx" structure, so that we only need to look at this and
the packet, no other memory when processing packets on the fast
RX path.

After doing the above per-packet checks, the data path collapses
down to a pretty simple conversion function taking advantage of
the data cached in the small fast_rx struct.

This should speed up the RX processing, and will make it easier
to reason about parallelizing RX (for which statistics will need
to be per-CPU still.)
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

On 32-bit platforms, the 64-bit counters we keep need to be protected
to be consistently read. Use the u64_stats_sync mechanism to do that.

In order to not end up with overly long lines, refactor the tidstats
assignments a bit.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When storing the last_rate_* values in the RX code, there's nothing
to guarantee consistency, so a concurrent reader could see, e.g.
last_rate_idx on the new value, but last_rate_flag still on the old,
getting completely bogus values in the end.

To fix this, I lifted the sta_stats_encode_rate() function from my
old rate statistics code, which encodes the entire rate data into a
single 16-bit value, avoiding the consistency issue.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Instead of touching the rx_stats.last_rx from the status path, introduce
and use a status_stats.last_ack variable. This will make rx_stats.last_rx
indicate when the last frame was received, making it available for real
"last_rx" and statistics gathering; statistics, when done per-CPU, will
need to figure out which place was updated last for those items where the
"last" value is exposed.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Move the averaged values out of rx_stats and into rx_stats_avg,
to cleanly split them out. The averaged ones cannot be supported
for parallel RX in a per-CPU fashion, while the other values can
be collected per CPU and then combined/selected when needed.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Legacy clients don't support P2P power save mechanism, and thus if a P2P GO
has a legacy client connected to it, it should disable P2P PS mechanisms.
Let the driver know about this with a new bss_conf parameter.
Signed-off-by: NAyala Beker <ayala.beker@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

By default, the rhashtable logic will fail to insert
objects if the key-chains are too long and un-balanced.

In the degenerate case where mac80211 is creating many
virtual interfaces connected to the same peer(s), this
case can happen.

St insecure_elasticity to true to allow chains to grow
as long as needed.
Signed-off-by: NBen Greear <greearb@candelatech.com>
[remove message, change commit message slightly]
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The original hand-implemented hash-table in mac80211 couldn't result
in insertion errors, and while converting to rhashtable I evidently
forgot to check the errors.

This surfaced now only because Ben is adding many identical keys and
that resulted in hidden insertion errors.

Cc: stable@vger.kernel.org
Fixes: 7bedd0cf ("mac80211: use rhashtable for station table")
Reported-by: NBen Greear <greearb@candelatech.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Some devices, like iwlwifi, have RSS queues. This may cause a
situation where a disassociation is handled in control path and
results in station removal while there are prior RX frames
that were still not processed in other queues. When they will
be processed the station will be gone, and the frames will be
dropped.
Add a synchronization interface to avoid that. When driver returns
from the synchronization mac80211 may remove the station.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

On error we jumped to the error label and returned the error code but we
missed releasing sinfo.

Fixes: 5fe74014172d ("mac80211: avoid excessive stack usage in sta_info")
Reviewed-by: NJulian Calaby <julian.calaby@gmail.com>
Signed-off-by: NSudip Mukherjee <sudip@vectorindia.org>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This will allow drivers to make more educated
decisions whether to defer transmission or not.

Relying on wake_tx_queue() call count implicitly
was not possible because it could be called
without queued frame count actually changing on
software tx aggregation start/stop code paths.

It was also not possible to know how long
byte-wise queue was without dequeueing.
Signed-off-by: NMichal Kazior <michal.kazior@tieto.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

When CONFIG_OPTIMIZE_INLINING is set, the sta_info_insert_finish
function consumes more stack than normally, exceeding the
1024 byte limit on ARM:

net/mac80211/sta_info.c: In function 'sta_info_insert_finish':
net/mac80211/sta_info.c:561:1: error: the frame size of 1080 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]

It turns out that there are two functions that put a 'struct station_info'
on the stack: __sta_info_destroy_part2 and sta_info_insert_finish, and
this structure alone requires up to 792 bytes.

Hoping that both are called rarely enough, this replaces the
on-stack structure with a dynamic allocation, which unfortunately
requires some suboptimal error handling for out-of-memory.

The __sta_info_destroy_part2 function is actually affected by the
stack usage twice because it calls cfg80211_del_sta_sinfo(), which
has another instance of struct station_info on its stack.
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Fixes: 98b62183 ("mac80211/cfg80211: add station events")
Fixes: 6f7a8d26 ("mac80211: send statistics with delete station event")
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The minimum chandef bandwidth calculation was done only in case
a new station was inserted (or when an existing station was removed).
However, it is possible that stations are inserted before they are
associated, e.g., when FULL_AP_CLIENT_STATE is supported and user
space adds stations unassociated.

Fix this by calling ieee80211_recalc_min_chandef() whenever
a station transitions in/out the associated state, and only
consider station marked as associated.
Signed-off-by: NIlan Peer <ilan.peer@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

My commit below broken PS-Poll handling. In case the driver
has no frames buffered, driver_release_tids will be 0, but
calling find_highest_prio_tid() with 0 as a parameter is
not a good idea:
fls(0) - 1 = -1.
This bug caused mac80211 to think that frames were buffered
in the driver which in turn was confused because mac80211
was asking to release frames that were not reported to
exist.
On iwlwifi, this led to the WARNING below:

WARNING: CPU: 0 PID: 11230 at drivers/net/wireless/intel/iwlwifi/mvm/sta.c:1733 iwl_mvm_sta_modify_sleep_tx_count+0x2af/0x320 [iwlmvm]()
ffffffffc0627c60 ffff8800069b7648 ffffffff81888913 0000000000000000
0000000000000000 ffff8800069b7688 ffffffff81089d6a ffff8800069b7678
0000000000000001 ffff88003b35abf0 ffff88000698b128 ffff8800069b76d4
Call Trace:
[<ffffffff81888913>] dump_stack+0x4c/0x65
[<ffffffff81089d6a>] warn_slowpath_common+0x8a/0xc0
[<ffffffff81089e5a>] warn_slowpath_null+0x1a/0x20
[<ffffffffc05f36bf>] iwl_mvm_sta_modify_sleep_tx_count+0x2af/0x320 [iwlmvm]
[<ffffffffc05dae41>] iwl_mvm_mac_release_buffered_frames+0x31/0x40 [iwlmvm]
[<ffffffffc045d8b6>] ieee80211_sta_ps_deliver_response+0x6e6/0xd80 [mac80211]
[<ffffffffc0461296>] ieee80211_sta_ps_deliver_poll_response+0x26/0x30 [mac80211]
[<ffffffffc048f743>] ieee80211_rx_handlers+0xa83/0x2900 [mac80211]
[<ffffffffc04917ad>] ieee80211_prepare_and_rx_handle+0x1ed/0xa70 [mac80211]
[<ffffffffc045e3d5>] ? sta_info_get_bss+0x5/0x4a0 [mac80211]
[<ffffffffc04925b6>] ieee80211_rx_napi+0x586/0xcd0 [mac80211]
[<ffffffffc05eaa3e>] iwl_mvm_rx_rx_mpdu+0x59e/0xc60 [iwlmvm]

Fixes: 0ead2510 ("mac80211: allow the driver to send EOSP when needed")
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

add ieee80211_iter_keys_rcu() to iterate over uploaded
keys in atomic context (when rcu is locked)

The station removal code removes the keys only after
calling synchronize_net(), so it's not safe to iterate
the keys at this point (and postponing the actual key
deletion with call_rcu() might result in some
badly-ordered ops calls).

Add a flag to indicate a station is being removed,
and skip the configured keys if it's set.
Signed-off-by: NEliad Peller <eliadx.peller@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This can happen when the driver needs to send less frames
than expected and then needs to close the SP.
Mac80211 still needs to set the more_data properly based
on its buffer state (ps_tx_buffer and buffered frames on
other TIDs).
To that end, refactor the code that delivers frames upon
uAPSD trigger frames to be able to get only the more_data
bit without actually delivering those frames in case the
driver is just asking to set a NDP with EOSP and MORE_DATA
bit properly set.
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Some devices or drivers cannot deal with having the same station
address for different virtual interfaces, say as a client to two
virtual AP interfaces. Rather than requiring each driver with a
limitation like that to enforce it, add a hardware flag for it.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Group station statistics by where they're (mostly) updated
(TX, RX and TX-status) and group them into sub-structs of
the struct sta_info.

Also rename the variables since the grouping now makes it
obvious where they belong.

This makes it easier to identify where the statistics are
updated in the code, and thus easier to think about them.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There's little point in keeping (and even sending to userspace)
the beacon_loss_count value per station, since it can only apply
to the AP on a managed-mode connection. Move the value to ifmgd,
advertise it only in managed mode, and remove it from ethtool as
it's available through better interfaces.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There's only a single caller of this function, so it can
be moved to the same file and made static.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The mac80211 code uses ktime_get_ts to measure the connected time.
As this uses monotonic time, it is y2038 safe on 32-bit systems,
but we still want to deprecate the use of 'timespec' because most
other users are broken.

This changes the code to use ktime_get_seconds() instead, which
avoids the timespec structure and is slightly more efficient.
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Cc: Johannes Berg <johannes@sipsolutions.net>
Cc: linux-wireless@vger.kernel.org
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Instead of using the out-of-line average calculation, use the new
DECLARE_EWMA() macro to declare a signal EWMA, and use that.

This actually *reduces* the code size slightly (on x86-64) while
also reducing the station info size by 80 bytes.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

According to 802.11-2012 13.3.1, a mesh STA should assign an AID
upon receipt of a mesh peering open frame rather than using the link
id of the peer.  Using the peer link id has two potential issues:
it may not be unique among the peers, and by its nature it is random,
so the TIM may not compress well.

In preparation for allocating it properly, use sta->sta.aid, but keep
the existing behavior of using the plid in the aid we send.
Signed-off-by: NBob Copeland <me@bobcopeland.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There are now a fairly large number of mesh fields that really
aren't needed in any other modes; move those into their own
structure and allocate them separately.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Currently, the station hash table lookup (or iteration) must
access two cachelines for each station - the one with the hash
table node, and the one with the MAC address.

Duplicate the MAC address next to the hash node to get rid of
this. Since the MAC address is static there's no consistency
problem introduced by this.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

As we're running out of hardware capability flags pretty quickly,
convert them to use the regular test_bit() style unsigned long
bitmaps.

This introduces a number of helper functions/macros to set and to
test the bits, along with new debugfs code.

The occurrences of an explicit __clear_bit() are intentional, the
drivers were never supposed to change their supported bits on the
fly. We should investigate changing this to be a per-frame flag.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The hashtable behaviour change was merged into the tree
at about the same time as the mac80211 use of rhashtable,
but of course these don't really conflict in the normal
sense. Enable hash table shrinking now.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Allow debug builds to configure the station hash table maximum
size in order to run with hash collisions in limited scenarios
such as hwsim testing. The default remains 0 which effectively
means no limit.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

My conversion of the mac80211 station hash table to rhashtable
completely broke the lookup in sta_info_get() as it no longer
took into account the virtual interface. Fix that.

Fixes: 7bedd0cf ("mac80211: use rhashtable for station table")
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

In order to speed up mac80211's TX path, add the "fast-xmit" cache
that will cache the data frame 802.11 header and other data to be
able to build the frame more quickly. This cache is rebuilt when
external triggers imply changes, but a lot of the checks done per
packet today are simplified away to the check for the cache.

There's also a more detailed description in the code.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Both minstrel (reported by Sven Eckelmann) and the iwlwifi rate
control aren't properly taking concurrency into account. It's
likely that the same is true for other rate control algorithms.

In the case of minstrel this manifests itself in crashes when an
update and other data access are run concurrently, for example
when the stations change bandwidth or similar. In iwlwifi, this
can cause firmware crashes.

Since fixing all rate control algorithms will be very difficult,
just provide locking for invocations. This protects the internal
data structures the algorithms maintain.

I've manipulated hostapd to test this, by having it change its
advertised bandwidth roughly ever 150ms. At the same time, I'm
running a flood ping between the client and the AP, which causes
this race of update vs. get_rate/status to easily happen on the
client. With this change, the system survives this test.
Reported-by: NSven Eckelmann <sven@open-mesh.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The mesh plink code uses sta->lock to serialize access to the
plink state fields between the peer link state machine and the
peer link timer.  Some paths (e.g. those involving
mps_qos_null_tx()) unfortunately hold this spinlock across
frame tx, which is soon to be disallowed.  Add a new spinlock
just for plink access.
Signed-off-by: NBob Copeland <me@bobcopeland.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

This allows drivers to request per-vif and per-sta-tid queues from which
they can pull frames. This makes it easier to keep the hardware queues
short, and to improve fairness between clients and vifs.

The task of scheduling packet transmission is left up to the driver -
queueing is controlled by mac80211. Drivers can only dequeue packets by
calling ieee80211_tx_dequeue. This makes it possible to add active queue
management later without changing drivers using this code.

This can also be used as a starting point to implement A-MSDU
aggregation in a way that does not add artificially induced latency.
Signed-off-by: NFelix Fietkau <nbd@openwrt.org>
[resolved minor context conflict, minor changes, endian annotations]
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

We currently have a hand-rolled table with 256 entries and are
using the last byte of the MAC address as the hash. This hash
is obviously very fast, but collisions are easily created and
we waste a lot of space in the common case of just connecting
as a client to an AP where we just have a single station. The
other common case of an AP is also suboptimal due to the size
of the hash table and the ease of causing collisions.

Convert all of this to use rhashtable with jhash, which gives
us the advantage of a far better hash function (with random
perturbation to avoid hash collision attacks) and of course
that the hash table grows and shrinks dynamically with chain
length, improving both cases above.

Use a specialised hash function (using jhash, but with fixed
length) to achieve better compiler optimisation as suggested
by Sergey Ryazanov.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

Instead of looking up the destination station twice in the TX path
(first to build the header, and then for control processing), save
it when building the header and use it later in the TX path.

To avoid having to look up the station in the many callers, allow
those to pass %NULL which keeps the existing lookup.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>