Ethernet-like decapping mode leaves IP protocol
frame not aligned to 4-byte boundaries. This leads
to re-aligning in mac80211 which in turn leads to
poor CPU cache behaviour on some machines.
Since HW doesn't allow to change payload offset
properly the solution is to force HW to decap in
Native Wifi mode which always has 24-bytes long
802.11 header (even for QoS frames). This means IP
frame is properly aligned in this decap mode.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
NWifi decap mode always reports 802.11 Data
Frames, even when QoS Data Frames are actually
received.
This made mac80211 not report frame priority
properly (since there was no QoS Control field).
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Simplify decapping code and make it easier to
understand.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Clarify how each decap mode works in one place.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
HW reports each A-MSDU subframe as a separate
sk_buff. It is impossible to configure it to
behave differently.
Until now ath10k was reconstructing A-MSDUs from
subframes which involved a lot of memory
operations. This proved to be a significant
contributor to degraded RX performance.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Instead of allocating sk_buff for a mere 16-byte
tx fragment list buffer use headroom of the
original msdu sk_buff.
This decreases CPU cache pressure and improves
performance.
Measured improvement on AP135 is 560mbps ->
590mbps of UDP TX briding traffic.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Use a saner goto scheme for failure handling. Also
group operations more sensibly.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Until now the all MSDU transfer related structures
were freed when all resources were unreferenced.
Now HTC transfer is freed independently and HTT
transfer is so too.
This yields a way more simpler ath10k_skb_cb and
should possibly enable parallel pipe processing
(which is now serialized in
ath10k_pci_process_ce routine).
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
This reduces number of memory accesses and
hopefully contributes to better performance in the
future.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
It's more efficient to simply check num_pending_tx
value instead of traversing whole bitmap of
msdu ids.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Commit e9bb0aa39 ("ath10k: delete struct ce_sendlist") broke
num_sends_allowed incrementing. num_sends_allowed
exceeded initial values and could overflow.
This code was supposed to replenish
num_sends_allowed for partial sendlist items (i.e.
before final sendlist item from a sendlist was
completed and could be processed by completion
handlers).
Fortunately it seems it did not cause any major breakage,
yet.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Commit be8b394390 ("ath10k: make WMI commands block by design") broke
the build if CONFIG_ATH10K_TRACING was enabled.
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
It's not really necessary to have this processed
in a worker. There are no sleepable calls (and
actually shouldn't be).
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
It is no longer used nor necessary since WMI
commands can block.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
The patch prevents beacon misses in some case of
heavy load on a system.
If a beacon can't be transmitted directly from an
SWBA event it will be left in arvif->beacon and
transmission will be retried once TX credits
become available.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
The patch removes HTC endpoint tx workers in
favour of direct command submission. This makes a
lot more sense for data path.
mac80211 queues are effectively stopped/woken up
in a more timely fashion preventing build up of
frames. It's possible to push more traffic than
the device/system is able to handle and have no
hiccups or performance degradation with UDP
traffic.
WMI commands will now report errors properly and
possibly block as they actively can wait for tx
credits to become available.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
This will be necessary for further changes in
command submission scheme.
Once HTC is cleaned up WMI commands will finally
block.
This requires for SWBA to be processed in a
non-atomic context for now. Once other necessary
changes are in this will be reverted.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
This will allow higher layers to anticipate and
act upon TX credits renewal. This will be
important for some future rework of WMI command
submission.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Credit calculation was overly complex
unnecessarily. Now skb dequeing is more unified.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
The driver core clears the driver data to NULL after device_release
or on probe failure. Thus, it is not needed to manually clear the
device driver data to NULL.
Signed-off-by: Jingoo Han <jg1.han@samsung.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
struct ce_sendlist is useless as we always add just one buffer onto it.
And most importantly, it's ugly as it doesn't use skb properly.
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
The peer PHY mode for 11ac operation needs to be determined
properly based on the channel bandwidth being used. Fix
this so that the proper mode is given to the firmware.
kvalo: earlier we used 11na-ht20 in STA mode for 11ac AP peer, this
patch changes that to 11ac-vht80. I didn't notice any change in
throughput in my tests, but nevertheless it's the right thing
to do.
Signed-off-by: Sujith Manoharan <c_manoha@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Just to unify with the rest of debug messages. Minimal functional changes,
only major ones are removal of the awkward "else" style in debug
messages.
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Fix them to follow the general logging style in ath10k.
While at it, add print id chip_id to the debug log.
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
core.c mostly deals with driver and firmware starting related
actions. And we can use the boot level also in other components,
like PCI and HTT.
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Otherwise if CONFIG_ATH10K_DEBUGFS won't be set we will
end up with multiple definitions and compilation failure
in each place the header is included.
Signed-off-by: Bartosz Markowski <bartosz.markowski@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
The double while() could end up running forever.
Inner while() would complete very fast. However
the completion processing could take enough time
for more completions to flow in. In that case the
outer while() would not terminate and run again,
and again. This could happen especially on a slow
host system.
This could lead to a system freeze during heavy
traffic. Note: this doesn't solve all known
starvation issues yet.
Signed-off-by: Michal Kazior <michal.kazior@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
There's no functional changes. Just a small cleanup.
Signed-off-by: Bartosz Markowski <bartosz.markowski@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
When configuring the host_interests over BMI, set the UART
baud rate to 19200. This is valid for QCA988X_2.0 devices.
kvalo: found during code review, there should not be any functionality
changes
Signed-off-by: Bartosz Markowski <bartosz.markowski@tieto.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
The indentation here implies this was meant to be
a multi-statement if, but it lacks the braces.
kvalo: add "ath10k: " prefix
Signed-off-by: Dave Jones <davej@fedoraproject.org>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
ath10k doesn't support qca988x hw1.0 boards anymore. Unfortunately
the PCI id is the same in hw1.0 and hw2.0 so ath10k tries to use
hw1.0 boards anyway. But without hw1.0 workarounds in place
ath10k just crashes horribly.
To avoid using hw1.0 boards at all add a chip id detection
and fail the probe if hw1.0 is detected:
[ 5265.786408] ath10k: ERROR: qca988x hw1.0 is not supported
[ 5265.786497] ath10k: Unsupported chip id 0x043200ff
[ 5265.786574] ath10k: could not register driver core (-95)
[ 5265.793191] ath10k_pci: probe of 0000:02:00.0 failed with error -95
Also add a warning if there's an unknown chip id but continue
the boot process normally anyway.
Reported-by: Zaki Bakar <zaki.bm@gmail.com>
Tested-by: Christian Lamparter <chunkeey@googlemail.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
ath10k_ce_init() and the functions it calls wakeup
the chip multiple times. Simplify that to call
ath10k_pci_wake() only once. This also makes it
easier to add error handling when wakeup fails.
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
We should not try to access hw if wakeup fails so add
proper error checking for that. Also add the timeout lenght
to the warning message.
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
