Protect all functions touching queue_stop_reasons by spin-lock, since
they are accessed by op_tx. Now there's no need to take the mutex
before caling wlcore_queue_xxx functions.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Start using the new hw_queue mechanism in mac80211 and give each AC in
each vif its own hw_queue number. This allows us to stop an AC in a vif
independently from other vifs.

Change the Tx watermark handling functions to count packets per AC in
vif. From now on fast links should not be able to hurt the throughput
of slow links on the same AC but on different vifs.

Change internal queue mgmt functions to operate per vif, to support the
new Tx watermark granularity. Make the global versions of the queue
stop/start functions to use the global mac80211 API for queue mgmt. This
helps in situations where the driver currently doesn't know all the vifs
that reside in mac80211. Recovery is a good example for such a case.

[Moved hw_base_queue addition into the wlcore_tx_get_mac80211_queue()
function; changed WARN_ONs to WARN_ON_ONCEs; simplified for loops;
fixed new checkpatch warnings. -- Luca]
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Only allow a PSM STA to congest FW memory when it is the single active
link. Being a single STA doesn't imply a single link - there might be
other links on other roles.

[Changed WARN_ON to WARN_ON_ONCE -- Luca]
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Update the 18xx FW status private part to include Tx related link
priorities. Introduce new HW ops to determine link priority per chip
family.

For 18xx the changes are:
- Suspended links are at most low priority and Tx for them is stopped
  beyond the suspend threshold.
- Active links now get their thresholds directly from FW
- There's a new "stop" threshold for active links, at which point a link
  stops receiving new packets.

Update the min 18xx FW version required to make sure suspended links
bitmap is advertised by the FW.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Require each incoming packet to have a valid vif. The injected Tx code
path was buggy (and unused), so disallow it altogether.

Cleanup a few places and add a warning so we can better discover
anomalies (corrupted skbs?) masquerading as injected Tx.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Prioritize EDCA by choosing the AC before anything else. Use the
fast/slow link bitmap in FW to improve the scheduling algorithm for
the multi-link scenario.

Set packet thresholds to determine if a given link is high or low
priority according to its speed. A slow link will be given high priority
if the amount of packets queued for it in the FW is lower than the
slow-threshold. Similarly, a fast link will be given high priority if
the number of its packets queued in FW is smaller than the high-threshold.

The improved algorithm:
1. Choose optimal AC according to FW utilization
2. Traversing the VIFs in a round-robin fashion, try to choose a high
   priority link. Links are traversed in a round-robin fashion inside a
   VIF.
3. If no high priority links are found, choose the first non-empty
   (low priority) link found in the round robin.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Move FW-allocation tracking code to the fw_status function and track
allocations made by all links. These will be incorporated in the
improved Tx scheduling algorithm.

Manually zero the system link counters on op_stop, as this link is not
allocated the normal way.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

This accounting will help find a vif that has data in a specific AC.
Otherwise we have to traverse all the links, which can be lengthy for
the AP case.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Make sure free_link is always called when removing an interface. This
ensures all skbs belonging to this interface are returned to mac80211.
Otherwise these dangling skbs might crash the system on the next
call to wl1271_tx_reset_link_queues(). This happens on recovery/stop or
an unsuccessful Tx flush.
Signed-off-by: NIdo Yariv <ido@wizery.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Sometimes the driver can perform a recovery while Tx is
active, this will trigger unnecessary warning which might
delay the recovery for more than 100 mS.
Signed-off-by: NVictor Goldenshtein <victorg@ti.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

The 12xx FW doesn't support non-zero session ids for AP-mode. Introduce
an appropriate quirk to make sure the session id is always zero when
needed.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

A new session_id is generated on link allocation.
it is saved in a global array and used later, on tx.

The new fw api adds new bcast/global_session_id
fields to start_role(ap) command, and a new session_id
field to add_peer command. align the driver with it.
Signed-off-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

The scan APIs of 12xx and 18xx are totally different.
Use some common functions as much as possible (e.g.
for setting scan channels), but split scan.c into
chip-specific scan.c files, each implementing its
own scan mechanism.

(in other words - move most of the current wlcore's
scan.c into wl12xx, and implement a similar mechanism
in 18xx, according to the new api)

New wlcore ops are introduced in order to call the
chip-specific scan functions.

The template indices used for each scan (regular/scheduled)
are also different between the chips, so set the correct
indices used for each scan type after identifying the chip.
Signed-off-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

implement the reamin_on_channel() callback by starting
a dev role (already associated with the current vif)
on the requested channel/band.

This channel is usually different from the channel
of the sta role, so pass it to wl12xx_roc() as well,
and notify mac80211 (async) when the fw is ready
on the new channel.

Now, in case of offchannel tx, we should use the dev
role hlid, instead of the sta hlid.
Signed-off-by: NEliad Peller <eliad@wizery.com>
Reviewed-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Make the connection flow simpler by starting
sta role on bssid change.

Currently, we start dev role when going idle-off,
and start the sta role only after association
indication. This complicates the connection
flow with some possible intermediate states.

Make it simpler by starting sta role on bssid change,
which now happens *before* auth req get sent.

Update the handling of mac80211's notifications
and change wl1271_join/unjoin accordingly -
* Split wl1271_join() into wlcore_join (tuning on
  a channel/bssid) and wlcore_set_assoc (configure
  sta after association).
* Rename wl1271_unjoin() to wlcore_unset_assoc(), as
  it is no longer the inversion of wl1271_join()
  (now it's only used to disconnect associated sta /
  joined ibss, without stopping the role).
* Set ssid before starting station role (needed for
  start_role(sta)

While on it, split wl1271_bss_info_changed_sta() into
some sub-functions.

since we no longer use dev role in the connection flow,
we now always use the hlid of the sta role.
Signed-off-by: NEliad Peller <eliad@wizery.com>
Reviewed-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Aggregation buffer size is set separately per 18xx/12xx chip family.
For 18xx aggragation buffer is set to 13 pages to utilize all
the available tx/rx descriptors for aggregation.

[Arik - remove redundant parts from the patch]
Signed-off-by: NIgal Chernobelsky <igalc@ti.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <luca@coelho.fi>

Force Tx during the flush if there are packets pending in the driver.
This actually solves a bug where we would get called from the mac80211
wq context, which would prevent tx_work from getting queued, even when
the mutex is unlocked.

Don't stop the queues needlessly if there's nothing to flush. Use a
larger delay when sleeping to give the driver a chance to flush and
avoid cpu busy looping. Re-arrange the loop so the last iteration is
not wasted.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <luca@coelho.fi>

Print how much time a flush took. This will help debug the time it takes
to switch between channels.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <luca@coelho.fi>

This solves interoperability issues with peer that don't seem to "hear"
management packets transmitted in higher rates. Based on a previous
patch by Igal Chernobelsky.

Cc: Igal Chernobelsky <igalc@ti.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <luca@coelho.fi>

When a function requests to recover, it would normally abort and will
not send any additional commands to the HW. However, other threads may
not be aware of the failure and could try to communicate with the HW
after a recovery was queued, but before the recovery work began.

Fix this by introducing an intermediate state which is set when recovery
is queued, and modify all state checks accordingly.
Signed-off-by: NIdo Yariv <ido@wizery.com>
Signed-off-by: NLuciano Coelho <luca@coelho.fi>

avoid using the skb CB for getting the appropriate sta.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Acked-by: NLuciano Coelho <coelho@ti.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

There was a bug hiding here since the hlid was sometimes inferred from
the sta, which might be invalid at this point.

Instead, propagate the hlid from the skb-queue where we got the skb
in the first place.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Acked-by: NLuciano Coelho <coelho@ti.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

After the latest mac80211 changes, the sta has
the ap's sta pointer even before association.

This cause the auth and assoc frames to be sent
with the standard ap's rates, rather than the
basic rates.

Change the tx rate policy logic to use the regular
ap rates only for data packets (so control and mgmt
packets will be sent with basic rates)
Signed-off-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Send EAPOLs using minimum basic rate for AP, STA, p2p GO and Client.
The patch fixes p2p connection issue with Realtek device in p2p
certification test 5.1.13 (DEVUT reinvokes Persistent Group).
Signed-off-by: NIgal Chernobelsky <igalc@ti.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

commit 14bba17b "wl12xx: Propagate errors from wl1271_raw_write32"
breaks down TX in certain scenarios. wl1271_irq_locked() propagates
errors from wl1271_tx_work_locked however it may return -EBUSY
when the FW queues are full which is a legitimate case and not a
a real error. In this case a recovery is triggered by wl1271_irq
and this keeps repeating itself so TX is completely broken.
Fix it by avoiding propagating return values as errors even if they
aren't. Only bus (SDIO or SPI) ops failures would be progagated
as only these should trigger recovery.
Signed-off-by: NEyal Shapira <eyal@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Propagate errors from wl1271_raw_write32 and request for recovery when
appropriate.
Also rename prefixes of wlcore functions which their prototypes had to
be changed.
Signed-off-by: NIdo Yariv <ido@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Propagate errors from wl1271_write and request for recovery when
appropriate.
Also rename prefixes of wlcore functions which their prototypes had to
be changed.
Signed-off-by: NIdo Yariv <ido@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Propagate errors from wl1271_read and request for recovery when
appropriate.
Also rename prefixes of wlcore functions which their prototypes had to
be changed.
Signed-off-by: NIdo Yariv <ido@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

EAPOLs are sent at high rates as they are considered
data packets. Some APs like Motorola Symbol AP7131 and AP650
don't respond well to these rates and don't respond with
EAPOL 3/4 consistently. When sending EAPOL 2/4 at 54Mbps
we've seen approx 30% success rate in getting EAPOL 3/4 response
while using 11Mbps we got 100% success.
To increase the chances of successful 4-Way handshake with
such APs, send EAPOLs with basic rate policy in order to avoid
high rates.
Signed-off-by: NEyal Shapira <eyal@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

In PG2 only the last frame in the aggregate buffer should be
aligned to the sdio block size. This frame's header msb should be
set to 0, while in all the previous frames in the aggregation
buffer, this bit should be set to 1.

[Add a HW op for setting the frame ctrl bit only for 18xx. Other minor
cleanups - Arik]

[Make the pre_pkt_send operation optional -- Luca]
Signed-off-by: NIdo Reis <idor@ti.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Adding casts of objects to the same type is unnecessary
and confusing for a human reader.

For example, this cast:

        int y;
        int *p = (int *)&y;

I used the coccinelle script below to find and remove these
unnecessary casts.  I manually removed the conversions this
script produces of casts with __force, __iomem and __user.

@@
type T;
T *p;
@@

-       (T *)p
+       p

Neatened the mwifiex_deauthenticate_infra function which
was doing odd things with array pointers and not using
is_zero_ether_addr.
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The 12xx set_key just calls the common wlcore_set_key function, in order
to program the keys into the FW.

The 18xx variant changes the spare block count when a GEM or TKIP
key is set. Also modify the get_spare_blocks HW op for 18xx to return
the correct numbers of spare blocks, according to what is currently
set in FW.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Stop network queues during Tx flush, and also drop other internal
mac80211 packets (mgmt) that may arrive when the queues are stopped.

When flush is done all driver queues are clear, forcefully if needed.

Protect the Tx flush operation with a new mutex, to prevent concurrency
that can mess us queue state.

Based on a patch by Eliad Peller <eliad@wizery.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Allow the driver to wake/stop the queues for multiple reasons. A queue
is started when no stop-reasons exist.

Convert all wake/stop queue calls to use the new API.

Before, a stopped queue was almost synonymous a high-watermark on Tx.
Remove a bit of code in wl12xx_tx_reset() that relied on it.

Internal packets arriving from mac80211 are also discarded when a queue
is stopped. A notable exception to this is the watermark reason, which
is a "soft"-stop reason. We allow traffic to gradually come to a halt,
but we don't mind spurious packets here and there. This is merely a flow
regulation mechanism.

Based on a similar patch by Eliad Peller <eliadWizery.com>.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Add a HW op for getting spare blocks.

12xx cards require 2 spare blocks for GEM encrypted SKBs, regardless
of VIFs or keys programmed into the FW.

18xx cards require 2 spare blocks when there are any connected TKIP or
GEM VIFs. For now always return 2 spare blocks, as this works with all
networks. The special case TKIP/GEM functionality is added at a later
patch.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

18xx chips do not require extra space in the TKIP header. Introduce a
new HW quirk to allow us to make this feature arch-specific. 12xx chip
will now have this quirk.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Multicast management frames (e.g. global deauth)
should be sent out on the bcast link, rather than
the global, which should be used only for pre-added
stations (e.g. for auth/assoc resp).
Signed-off-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NEyal Shapira <eyal@wizery.com>
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Parse the peer ht_cap element containing MCS8-MCS15 rates and pass it
to the FW. Rates unsupported by the HW will be sanitized by mac80211
before reaching us.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Some chip families are capable of checksumming certain classes of Tx
packets in HW. Indicate this fact in the netdev features and perform the
HW checksum by protocol type for the 18xx family.

Fix the location of the skb network header when we move it so we can
rely on it when setting the checksum.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>

Implement immediate Tx completion for the 18xx family. Move 18xx
specific Tx code to new tx.c/h files and create helper header files
for definitions.
Signed-off-by: NArik Nemtsov <arik@wizery.com>
Signed-off-by: NLuciano Coelho <coelho@ti.com>