txdata->enabled is used in order to determine if the TX vring
is valid. As the data transmit is handled in a different context,
in case txdata->enabled is set before vring->hwtail is updated,
an old or corrupted vring->hwtail can be used.
Protect setting of txdata->enabled and vring->hwtail to prevent a
case where TX vring start handling TX packets before setting
vring->hwtail.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
In case we fail to map one of the TSO SKB fragments, we need to
clear all the mapped descriptors, from swhead to swhead+descs_used-1.
Change the desc index calculation to
i = (swhead + descs_used - 1) % vring->size;
to prevent unmpping of (swhead + descs_used) descriptor that wasn't
mapped.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
napi_synchronize is called before releasing the vring, with the
assumption that setting txdata->enabled to 0 will prevent handling
of this vring in the next scheduled napi.
To guarantee this assumption, a memory barrier is added after disabling
the txdata.
In addition, as the ctx is zeroed in wil_tx_complete after this
descriptor is handled (protected by wmb), ctx needs to be checked
before releasing this descriptor in wil_vring_free.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
add memory barrier after allocating new rx descriptors, before
updating the hwtail.
This will guarantee that all writes to descriptors (shared memory)
are done before committing them to HW.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
There are 2 possible race conditions, both are solved by addition of
memory barrier:
1. wil_tx_complete reads the swhead to determine if the vring is
empty. In case the swhead was updated before the descriptor update
was performed in __wil_tx_vring/__wil_tx_vring_tso, the completion
loop will not end and as the DU bit may still be set from a previous
run, this skb can be handled as completed before it was sent, which
will lead to double free of the same SKB.
2. __wil_tx_vring/__wil_tx_vring_tso calculate the number of available
descriptors according to the swtail. In case the swtail is updated
before memset of ctx to zero is completed, we can handle this
descriptor while later on ctx is zeroed.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Network stack can try to transmit data while AP / STA is
disconnected.
Change this print-out to debug level as this should not be
handled as error.
This patch also adds wil_dbg_ratelimited, used to limit the
above print-out.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Switch to auto-generated version of wmi.h which is maintained
by FW team. This will allow better sync between teams in the
future and avoid bugs because of unexpected API changes.
The wmi.h will have many differences but most are cosmetic.
It also includes these real differences:
1. is_go parameter added to BCON_CTRL and START_PCP commands.
2. max_rx_pl_per_desc added to CFG_RX_CHAIN command.
3. various small API updates that are not currently used by
driver.
Signed-off-by: Lior David <qca_liord@qca.qualcomm.com>
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
wil_tx_vring locks the vring_tx_data lock before accessing the TX
vring to check if it is enabled and valid for use.
In case of quick disconnect / connect events for the same station,
spin_lock(&txdata->lock) can be called during the lock initialization
in the vring init function.
To prevent such a race, the TX vrings spin lock should be initialized
once during wil6210 driver initialization.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
In the current solution wil6210 configures the vring in a worker
and holds only one pending CID. This implementation may lead to
race conditions between connect and disconnect events of multiple
stations or fast connect/disconnect events of the same station.
In order to allow the removal of the connect worker and handling of
WMI_VRING_CFG_DONE_EVENTID in the connect event, the WMI replies
that provide the reply in a given buffer needs to be handled
immediately in the WMI event interrupt thread.
To prevent deadlocks, WMI replies that requires additional
handling are still handled via the events list.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
In TX vring allocation, vring_tx->va is allocated before WMI command to
configure the vring is sent to the device. As the WMI command can take
time to complete, it can lead to scenarios where vring_tx->va is not NULL
but vring is still not enabled.
This patch adds a check that vring_tx_data is enabled before returning
a valid TX vring.
Signed-off-by: Maya Erez <qca_merez@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
When Tx ring full is encountered with TSO,
printout of "DMA error" was wrongly printed.
In addition, in case of Tx ring full return
proper error code so that NETDEV_TX_BUSY is
returned to network stack in order not to
drop the packets and retry transmission of the
packets when ring is emptied.
Signed-off-by: Hamad Kadmany <qca_hkadmany@qca.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Situations observed when IP stack schedules lots of
frames for Tx while no connection (connection lost,
for example). In this case, dmesg bloated with error
message "FW not connected", printed for every frame.
Ratelimit this error message to avoid dmesg pollution.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
BAR frames delivered to the host via Rx path; whole BAR frame
get delivered. Advance sequence in the reorder buffer and release
old frames, as per IEEE802.11 spec.
Firmware will reply to BAR, driver responsibility is only reorder
buffer management.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
On Rx, when invalid frame is received and dropped,
reaping of next frames from Rx ring is stopped.
This stops NAPI polling and re-enables the Rx interrupt.
However, in cases where no more frames received,
interrupt will not be triggered and rest of Rx frames
will not be processed.
Skip bad frames and continue to reap Rx packets when
such frames are encountered, and add statistics for
such frames for debug.
Signed-off-by: Hamad Kadmany <qca_hkadmany@qca.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Allow network stack part of Rx processing to run on separate core,
relaxing CPU utilization on the core used for Rx NAPI.
If RXHASH feature is enabled, the driver sets rxhash of each skb to 1
to enable RPS. The core for processing the rx skb is determined by RPS
mechanism according to rx_cpus bit mask which is configured at user level.
For processing skbs on different core from the core which processes
the interrupts, it is recommended not to enable core 0 in rx_cpus bit mask.
Signed-off-by: Vladimir Shulman <QCA_shulmanv@QCA.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Replace macros like "R", "W", "S", "C", defined multiple times,
with inline functions "wil_[rwsc]".
Use "readl" and "writel" instead of "ioread32" and "iowrite32"
since it is granted that memory transactions are used,
not port ones like IN/OUT
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Driver report supported TSO (v4 & v6) and IP checksum offload
in addition to previously supported features. In data path
skbs are checked for non-zero gso_size, and when detected sent
to additional function for processing TSO SKBs. Since HW does not
fully support TSO, additional effort is required from the driver.
Driver partitions the data into mss sized descriptors which are
then DMAed to the HW.
Signed-off-by: Vladimir Shulman <QCA_shulmanv@QCA.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Provide detailed statistics for the Rx frames per MCS
Statistics printed in "stations" debugfs entry
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Introduce 2 types of GTK, Tx (for this STA) and Rx (for each peer).
Now, AP has only Tx GTK, STA - only Rx one. PBSS not supported yet;
for it, continue using pseudo-DMS.
Handle per-vring .1x state, update it from WMI_VRING_EN_EVENTID
event. This allows unification for unicast and broadcast vrings.
This mechanism replaces former per-CID "data_port_open"
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Rx descriptor fields accordingly to the updated
hardware documentation
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
Multicast is not yet properly supported for all connection types;
actually, only non-secure AP is supported. For all other cases,
fall back to old "pseudo-DMS" approach. Namely, for:
- PBSS
- secure connection
When re-routing MCAST Rx->Tx on the AP, do not Tx back to the origin
of the frame
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Use dedicated vring for multicast frames; this vring allocated for
AP and PBSS (both P2P GO and client) configurations
For short frames, use MCS0; for long - MCS1
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
For the networking code and for hardware network accelerators,
it is better to have IP header 4*n aligned. On the other side,
DMA on Rx path require buffer to be aligned on 4*n as well.
Having 14 bytes of Ethernet header, these 2 alignment
requests are in contradiction.
To solve this, order hardware offload block to not remove
SNAP header. This adds extra 6 bytes between addresses and
ethertype, making it 20 bytes total. This way, both buffer and
IP header are 4*n aligned. Remaining is only to remove SNAP
by shifting addresses 6 bytes. This involves data copying, so
this feature should be disabled unless required by the platform.
Module parameter "rx_align_2" (bool, default - false)
introduced to control this feature. Feature is completely disabled
when parameter is false.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Use temporal variable for often used vring->swhead;
and use proper index in debug printing - vring->swhead
used before was modified in wil_vring_advance_head
and then increased value was used in debug print
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
For the AP, configuration may say not to bridge traffic between
wireless clients. This is conveyed from user space (ex: hostapd has
ap_isolate parameter) with NL80211_CMD_SET_BSS, to the driver's
cfg80211 ops method change_bss
Add support for this setting.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
This is for AP only. If Rx data frame targeted to one of associated clients,
transmit it back to the wireless media and don't deliver to the host.
For the multicast frames, deliver to both host and wireless media.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Change threshold to be variable debugfs entry from hard-coded 0.
Default threshold value is 16 descriptors because HW is capable
of fetching up to 16 descriptors at once.
Signed-off-by: Vladimir Shulman <qca_shulmanv@qca.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Mark expected branches using likely()/unlikely().
Do it on high performance route - data path and interrupts
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
In the old hardware, bug existed that caused DA and SA for every
Rx packet to be swapped in the AP mode.
New hardware has fix for this bug. Enable this fix in the
hardware.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Print Tx descriptors
Print ring index for all Tx related messages
Sort prefixes: Tx for transmit, TxC for completion, added "D"
for descriptor related prints
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Use spinlock, this should not impact Tx as lock is always free
except for de-allocation.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Non-cacheable Tx descriptor for the last fragment of
multi-fragment frame used to be written back twice.
Refactor code to always write non-cacheable descriptor
only once
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
In the Tx, work around used to force destination index 0
to be used. This is no more necessary, as firmware supports
multiple destinations
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Implement Tx status reporting using skb_complete_wifi_ack().
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Sync documentation for the Tx/Rx descriptors with the
firmware/hardware documentation.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
In the reordering block, Ethernet DA was checked for MCAST, this is wrong.
Check instead MCAST indication from 802.11 MAC header. Hardware saves
this into Rx descriptor.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
In STA mode, all Tx should be directed to the same VRING towards the AP.
Thus, look up for the 1-st eligible VRING and use it.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
wil->status used as bitmap; use DECLARE_BITMAP for it.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
When configuring Tx/Rx VRING's, driver need to specify max. MPDU size
It should take into account all overhead introduced by 802.3->208.11
transformation.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
For manual ADDBA configuration, allow to set desired window size or
disable automatic mechanism.
Introduce module parameter (int) agg_wsize. It can be changed on run time,
will be taken into account on the next connect. Interpretation:
- <0 - disable automatic ADDBA; intended for manual testing through debugfs
- 0 - use automatically calculated window size
- >0 - use this for window size. Clipped by maximum supported by the hardware
with current environment.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Upon Tx vring creation, initiate BACK establishment
with maximum possible window size.
When establishing secure connection, there is EAPOL data exchange
between connection itself and "data port open", where security
is done and non-EAPOL data may be transferred. It is better to
send EAPOL frames using normal ACK because of firmware considerations.
send ADDBA only is 2 conditions met:
- data port open for the corresponded STA
- vring created
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Introduce BACK establishment procedures; decision logic is not implemented
yet; debugfs entry 'addba' used to manually trigger addba/delba for ringid 0.
debugfs usage:
to establish BACK with agg_wsize 16:
echo 16 > /sys/kernel/debug/ieee80211/phy0/wil6210/addba
to delete BACK:
echo 0 > /sys/kernel/debug/ieee80211/phy0/wil6210/addba
to change agg_wsize, one need to delete BACK and establish it anew
ADDBA flow for:
- originator
Tx side (initiator) sends WMI_VRING_BA_EN_CMDID providing
agg_wsize and timeout parameters.
Eventually, it gets event confirming BACK agreement - WMI_BA_STATUS_EVENTID
with negotiated parameters. On this event, update Tx vring data
(struct vring_tx_data) and display BACK parameters on debugfs
- recipient
Rx side (recipient) firmware informs driver about ADDBA with
WMI_RCP_ADDBA_REQ_EVENTID, driver process it in service work
queue wq_service. It adjusts parameters and sends response
with WMI_RCP_ADDBA_RESP_CMDID, and final confirmation provided
by firmware with WMI_ADDBA_RESP_SENT_EVENTID. In case of success,
driver updates Rx BACK reorder buffer.
policy for BACK parameters:
- aggregation size (agg_wsize * MPDUsize)) to not exceed 64Kbytes
DELBA flow for:
- originator
driver decides to terminate BACK, it sends WMI_VRING_BA_DIS_CMDID
and updates struct vring_tx_data associated with vring; ignore
WMI_DELBA_EVENTID.
- recipient
firmware informs driver with WMI_DELBA_EVENTID,
driver deletes correspondent reorder buffer
ADDBA request processing requires sending WMI command, therefore
it is processed in work queue context. Same work queue used as for
connect, it get renamed to wq_service
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Hardware doesn't place any restrictions on the buffer alignment,
consider this TODO resolved.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
In the wil_tx_ring, error message printed when tx attempted
while vring has no space to accommodate all fragments of frame.
Normally, such situation handled by stopping tx queue.
But, if tx queue is by-passed (like pktgen does), this error
will be triggered at high rate and dmesg will be flooded with
this message. Whole system may become unstable and hang with
no possible recover except power cycle.
Rate-limit it to prevent dmesg flooding.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
Allow to configure VRING size for both Rx and Tx via module parameters:
rx_ring_order and tx_ring_order. Parameters are ring size orders, i.e.
ring size calculated as 1 << order.
Defaults for both Tx and Rx are order 9, i.e. size 512
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
It is possible to configure driver using mtu_max module parameter
by setting it to value in range of 68..7920 inclusive.
This is sub-optimal performance-wise in case packet is larger than 1 page.
mtu_max default value is 2228.
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
HW supports upto 2304 packet size on the air.
HW is responsible for adding (Tx) or removing (Rx) the following headers:
802.11 hdr: 26B
SNAP: 8B
CRC: 4B
Security (optional): 24B
HW adds max 62B to the payload passed from driver. It means driver can use
max packet size of 2304-62 = 2242B
Signed-off-by: Dedy Lansky <qca_dlansky@qca.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
added misc printouts in some init/uninit functions for better traceability
Signed-off-by: Dedy Lansky <qca_dlansky@qca.qualcomm.com>
Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
Maya Erez