use proper format %pad for the dma_addr_t arguments;
prefix %p with 0x, as %p don't print is by itself
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Provide clear definition of the watermarks for the vring descriptor space.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When disconnecting some CID, corresponded Tx vring get released. During vring
release, all descriptors get freed. It is possible that Tx NAPI working on the same
vring simultaneously. If it happens, descriptor may be double freed.

To protect from the race above, make sure NAPI won't process the same vring.
Introduce 'enabled' flag in the struct vring_tx_data. Proceed with Tx NAPI only if
'enabled' flag set. Prior to Tx vring release, clear this flag and make sure NAPI
get synchronized.

NAPI enablement status protected by wil->mutex, add protection where it was
missing and check for it.

During reset, disconnect all peers first, then proceed with the Rx vring. It allows for
the disconnect flow to observe proper 'wil->status' and correctly notify cfg80211 about
connection status change
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

GRO is easy to enable when already using NAPI framework,
and it improves CPU utilisation. Enable it by default.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When switching between STA and AP modes, memory allocated for Rx vring leaks
This is because start_ap() allocates Rx vring but stop_ap() do not free it.
Logically, Rx vring is not valid (HW can't use it anymore), so free it in reset()
Also, check double init for Rx vring and bail out with -EINVAL
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When FW crashes, dmesg get polluted with the "FW not ready"
error message. Print it only once per FW lifecycle
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Introduce enum to describe mapping type; allow 'none' in addition to
'single' and 'page'; this is preparation for GSO
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When setting fragmented skb for Tx, assign skb to the last descriptor
and set number of fragments in the 1-st one
On Tx complete, HW sets "DU" bit in Tx descriptor only for the last
descriptor; so search for it using number of fragments field.
Middle descriptors may have "DU" bit not set by the hardware.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When connection established, as reported by WMI_CONNECT_EVENTID,
4-way handshaking required for the secure connection is not done
yet. It is indicated by another WMI event. Wait for it and only then
allow data traffic. In case of non-secure connection, FW reports
"data port open" immediately after connection.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Encode/decode helpers
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Groupcast frames are not subject for BACK reordering because they are not
ACK'ed and one can't request re-transmitt
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Need to fill translation table before calling WMI with WMI_VRING_CFG_CMDID
since firmware may generate events during this call; and events need translation
table filled to be properly dispatched
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Calculate statistics per connection, report with "iw station dump"
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Hardware do not support "real" broadcast on the air.
Use method similar to the Directed Multicast Service (DMS) as
described in the 10.23.15.2 "DMS procedures"

This service copies frame and delivers unicast for each associated peer

Do the following: send original frame to 1-st Tx vring, and send copies
to all other active vrings.

As currently hardware/firmware don't support A-MSDU, convert broadcast
frame to unicast instead of wrapping it in A-MSDU
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When running multiple connections, hardware can't do BACK reordering
and it should be done on the host.

Model after mac80211's implementation. Drop RCU for now;
to be re-added when BACK will be stabilized

BACK handshaking is not implemented yet in the hardware,
pretend it was done to support the way FW operating
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Enable multiple (up to 8 - HW/FW limitation) simultaneous connections.
Each connection has its own CID (connection ID) that describes chip's
beam-forming entity. Tx Vring should refer to correct CID for frame to reach
its destination.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This fixes bug found by the kbuild test robot:

tree:   git://git.kernel.org/pub/scm/linux/kernel/git/linville/wireless-next.git master
head:   1e2f9295
commit: 1cbbcb08 [135/140] wil6210: prefetch head of packet
config: make ARCH=microblaze allyesconfig

All error/warnings:

   drivers/net/wireless/ath/wil6210/txrx.c: In function 'wil_vring_reap_rx':
>> drivers/net/wireless/ath/wil6210/txrx.c:381:2: error: implicit declaration of function 'prefetch' [-Werror=implicit-function-declaration]
     prefetch(skb->data);
     ^
   cc1: some warnings being treated as errors

vim +/prefetch +381 drivers/net/wireless/ath/wil6210/txrx.c

   375                  wil_err(wil, "Rx size too large: %d bytes!\n", dmalen);
   376                  kfree_skb(skb);
   377                  return NULL;
   378          }
   379          skb_trim(skb, dmalen);
   380
 > 381          prefetch(skb->data);
   382
   383          wil_hex_dump_txrx("Rx ", DUMP_PREFIX_OFFSET, 16, 1,
   384                            skb->data, skb_headlen(skb), false);
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

As soon as skb is ready to be reaped, prefetch 1-st cache line.
This accelerates data access that is performed later, during the
packet classification by the driver and IP stack.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Bit DMA_CFG_DESC_TX_OFFLOAD_CFG_L3T_IPV4_POS should be set for IPv4
only. Don't set it for IPv6
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Fix for TCP iperf from Windows to Linux stall after about 1sec

Hardware reports false errors in some situations:

Microsoft IP stack, in violation of RFC 1624, set TCP checksum that should be 0x0
as 0xffff. hardware report Rx csum error. If HW csum absolutely trusted,
this frame can be never received, as re-transmitted one will have same csum problem.
In addition, it mess up block ack reorder buffer, as if packet dropped, it is not score boarded
there.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Add support for TCP and UDP HW checksum offloading.
RX chain is allways configured for offload mode.
In case of checksum error in RX path the DMA L4 error bit(5)
will be set to 1 and driver will drop the packet.
TX checksum offloading is configrable (ethtool -K).
TX descriptors are configured for checksum offload according
to the SKB protocol type (TCP/UDP, IPV4/6), Upon mismatch drop
the TX packet (checksum required but not TCP/UDP IPV4/6 type).
Signed-off-by: NKirshenbaum Erez <erezk@wilocity.com>
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Release fragments in the order of allocation; including one for skb head
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Finish all SW context modifications prior to notifying hardware

It used to be race condition: if HW finish Tx and issue Tx completion IRQ very fast,
prior to SW context update in wil_tx_vring, Tx completion will mis-handle descriptor, as
SW part will have no skb pointer stored.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Enable adding more data to the SW context.
For now, add flag "mapped_as_page", to separate decisions on free-ing skb
and type of DMA mapping.
This allows linking skb itself to any descriptor of fragmented skb.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

No need to modify HW descriptor, as it will be re-initialized on Tx.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When destroying Rx vring, branch for Rx used wrong Tx descriptor:
while SW context was taken for "head", HW descriptor was, by mistake,
taken from "tail"
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Map BIT 9 in TX DMA DWARD 0 as HW write back option.
We must turn on this option in the last TX descriptor,
this is required for old HW compatability.
This option indicate to HW that WB is required for this descriptor.
Signed-off-by: NKirshenbaum Erez <erezk@wilocity.com>
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

The vring index (MAC queue id) must be set in all TX descriptors
otherwise HW will fail to release descriptors for a specific vring
(disconnect or vring switch flows).
This is normally occurs when fragmentation required, if vring index
will not be the same for all SKB descriptors HW will fail to flush
this MAC queue.
Signed-off-by: NKirshenbaum Erez <erezk@wilocity.com>
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

No more need for special processing of EAPOL, FW can now send EAPOL frames
using normal Tx queue for TID 0

This fixes "schedule while atomic" bug - start_xmit called in softirq context;
while WMI mechanism that was used may sleep.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Use kfree_skb() instead of kfree() to free sk_buff.

Introduced by commit e270045b
(wil6210: Sanity check for reported DMA length)
Signed-off-by: NWei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Packet drop may be caused by various flows, like disconnect
while Tx packets was queued; this should not lead to stopping
of the Tx queue, or all Tx get stalled.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

- Introduce common code for Tx/Rx descriptor physical address set/parse
- Fix endianness for address fields
- consistent descriptor naming: '_d' for non-cached memory, 'd' for cached copy
- wil_tx_desc_map now modify cached copy, no need for 'volatile'
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Introduce NAPI for Rx and Tx completion.

This fixes packet reordering that happens when Rx handled right in
the IRQ: netif_rx puts packet in 'percpu' queue, then network stack
fetches packets from 'percpu' queues for processing, with different
pattern of queue switching. As result, network stack see packets
in different order. This causes hard to understand TCP throughput
degradation in about 30min

Complete polling if only one packet was processed - this eliminates
empty polls that would be otherwise done at the end of each burst
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Trace the following:
- WMI cmd/event
- log events
- interrupts
- Tx/Rx
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Move packet dump to the earliest location where it is known to have
valid data.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

If Rx descriptor contains garbage, it is possible to access memory beyond
allocated buffer.

Check this condition and drop Rx if reported length is
unreasonable large
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Hardware uses little endian for the Tx/Rx descriptors field 'length',
do appropriate conversions
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Original Tx descriptor stored is in non-cached area for DMA;
copy it to the cached memory to speed-up access
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

vring size is known from the beginning, fill it immediately
in the struct initializer
This is minor optimization that reduces code size.
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Rx descriptors stored in non-cacheable memory area for DMA.
Non-cacheable memory causes long access time from CPU.

Copy rx descriptor to the skb->cb, and use this copy.
It provides faster memory access, and will be usefull to keep
Rx information for later processing (BACK reorder)
Signed-off-by: NVladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>