Now that AQL works well on all mt76 drivers, completely replace the arbitrary
burst sizing and number of bursts logic for tx scheduling.
For the short period of time in which AQL does not work well yet, limit each
stations to 16 in-flight packets that have no estimated tx time.
This should avoid filling the queue if a station connects and queues up a
large number of packets before rate control information is available, especially
with hardware rate control
Signed-off-by: Felix Fietkau <[email protected]>
On 7615 and newer, DMA completion only triggers unmap, but not free of queued
skbs, since pointers to packets are queued internally.
Because of that, there is no need to process the main data queue immediately
on DMA completion.
To improve performance, mask out the DMA data queue completion interrupt and
process the queue only when we receive a txfree event.
This brings the number of interrupts under load down to a small fraction.
Signed-off-by: Felix Fietkau <[email protected]>
In order to avoid keeping work like tx scheduling pinned to the CPU it was
scheduled from, it makes sense to switch from tasklets to kernel threads.
Signed-off-by: Felix Fietkau <[email protected]>
On 7615 and newer, DMA completion only triggers unmap, but not free of queued
skbs, since pointers to packets are queued internally.
Because of that, there is no need to process the main data queue immediately
on DMA completion.
To improve performance, mask out the DMA data queue completion interrupt and
process the queue only when we receive a txfree event.
This brings the number of interrupts under load down to a small fraction.
Signed-off-by: Felix Fietkau <[email protected]>
The previous scheduling round may have been limited by AQL.
More frames might be available after the tx free run.
Signed-off-by: Felix Fietkau <[email protected]>
With threaded NAPI support it becomes more likely that tx cleanup and tx
enqueuing will run in parallel. In order to avoid queue starvation issues
under load, update q->queued immediately.
Signed-off-by: Felix Fietkau <[email protected]>
With threaded NAPI, the rx handler function is no longer bound to the CPU that
fired the interrupt, which significantly helps to spread the workload over
multiple CPUs, especially when multiple devices are using threaded NAPI at the
same time.
Signed-off-by: Felix Fietkau <[email protected]>
Since DMA completion does not imply tx completion, it makes more sense to
poll for airtime from mt7915_mac_tx_free.
Reduce the runtime of the function by moving all items from dev->sta_poll_list
to a local list once and process any stations that were added afterwards
on the next run
Signed-off-by: Felix Fietkau <[email protected]>
- change the WTBL LMAC access function to set the mapping window only once
- use ac * 2 as offset, since each AC has separate words for rx and tx
Signed-off-by: Felix Fietkau <[email protected]>
Using dev_kfree_skb for tx skbs breaks AQL. This worked until now only
by accident, because a mac80211 issue breaks AQL on drivers with firmware
rate control that report the rate via ieee80211_tx_status_ext as struct
rate_info.
Signed-off-by: Felix Fietkau <[email protected]>
The implementation of embedding WTBL update inside the STA_REC update is buggy
on the MT7615 v2 firmware. This leads to connection issues after a station has
connected and disconnected again.
Switch to the v1 MCU API ops, since they have received much more testing and
should be more stable.
On MT7622 and later, the v2 API is more actively used, so we should keep using
it as well.
Fixes: 6849e29ed92e ("mt76: mt7615: add starec operating flow for firmware v2")
Cc: [email protected]
Signed-off-by: Felix Fietkau <[email protected]>
Use the txwi data as primary source information to avoid touching skb data
Use bitfield instead of state variable + spinlock
Signed-off-by: Felix Fietkau <[email protected]>
mt7615 and newer drivers do not need this, since they use sequence number offload
Moving this code also reduces the number of callsites to make it easier to review
Signed-off-by: Felix Fietkau <[email protected]>
Queueing a per-sta work item from the tx free path can become very expensive
under load. This work is only supposed to pull rate control stats every
second and deal with rate control changes.
Additionally, the rate control update code was wrong, because it was
confusing bit masks and bit numbers in test_bit.
Fix this by introducing a dedicated device work item for rate control
updates, and by polling station stats from the phy mac work.
Stations requiring polling or rate control updates are added to lists
protected by dev->sta_poll_lock.
Signed-off-by: Felix Fietkau <[email protected]>
Only clear unmasked interrupts. If an interrupt is temporarily masked,
its pending events need to be processed later, even if another interrupt
happened in the mean time.
Disable interrupts in one call before scheduling
Signed-off-by: Felix Fietkau <[email protected]>
If an interrupt is temporarily masked, its pending events need to be processed
later, even if another interrupt happened in the mean time.
Signed-off-by: Felix Fietkau <[email protected]>
Only clear unmasked interrupts. If an interrupt is temporarily masked,
its pending events need to be processed later, even if another interrupt
happened in the mean time.
Signed-off-by: Felix Fietkau <[email protected]>
A small part of the first skb buffer is passed to the firmware for parsing
via DMA, while the full buffer is passed as part of the TXP.
Avoid calling DMA unmap on the first part (with a different length than map)
Signed-off-by: Felix Fietkau <[email protected]>
Supports sending a configurable number of packets with a specific rate
and configurable tx power levels / antenna settings, as well as displaying
rx statistics.
Signed-off-by: Felix Fietkau <[email protected]>
This can be used for calibration in the manufacturing process.
It supports sending a configurable number of packets with a specific rate
and configurable tx power levels / antenna settings.
It also supports receiving packets and showing some statistics, including
packet counters and detailed RSSI information.
It will only be compiled in if CONFIG_NL80211_TESTMODE is enabled
Signed-off-by: Felix Fietkau <[email protected]>
All drivers use this in pretty much the same way. Moving it to core helps with
some checks for the upcoming testmode support
Signed-off-by: Felix Fietkau <[email protected]>
Now that the bus access functions can use mapping for accessing full
register addresses, use it for WF_PHY registers to keep them constant.
Needed for follow-up work on testmode support
Signed-off-by: Felix Fietkau <[email protected]>
Unlike on earlier chips, DMA completion on MT7615 does not imply actually
having sent out any packets.
Since AQL will prevent filling the hardware queues and will only allow more
packets to be passed to the driver after tx completion, it makes much more
sense to schedule the tx tasklet there.
This is also needed for scheduling tx in testmode support
Signed-off-by: Felix Fietkau <[email protected]>
Avoid adding MT7615_EEPROM_SIZE twice.
Rename MT7615_EEPROM_EXTRA_DATA to MT7615_EEPROM_FULL_SIZE, since it already
includes MT7615_EEPROM_SIZE
Fixes: ad380ad1ebbe ("mt76: mt7615: add support for applying DC offset calibration from EEPROM")
Signed-off-by: Felix Fietkau <[email protected]>
Limits are used to update the channel max_power settings and also passed
to the firmware on channel changes
Signed-off-by: Felix Fietkau <[email protected]>
This subnode can be used to set per-rate tx power limits either per
country code / regdomain or globally.
These limits are typically provided by the device manufacturers and are
used to limit sideband emissions and stay within regulatory limits
Signed-off-by: Felix Fietkau <[email protected]>
Fixes a crash on MMIO devices when running into the watchdog reset
Fixes: d3377b78cec6 ("mt76: add HE phy modes and hardware queue")
Signed-off-by: Felix Fietkau <[email protected]>
mt76 core uses ffs() to find the next free bit. This works well for 32 bit
architectures where BITS_PER_LONG is 32. ffs only checks 32 bit values, so
allocation fails on 64 bit architectures.
Additionally, the wcid mask array was too small in cases where the array
was not a multiple of BITS_PER_LONG.
Fix this by making the wcid mask array u32 instead and use DIV_ROUND_UP
for the size, just in case we ever bump it to a value that's not a multiple
of 32.
Signed-off-by: Felix Fietkau <[email protected]>
All drivers before MT7915 have a limit of 128 WCID entries. Stop relying
on ARRAY_SIZE(dev->mt76.wcid), since it no longer reflects that limit.
Fixes: 49e649c3e0a6 ("mt76: adjust wcid size to support new 802.11ax generation")
Reported-by: kbuild test robot <[email protected]>
Signed-off-by: Felix Fietkau <[email protected]>
On top of the EEPROM target power, each rate can also has a power offset.
On many devices, this power offset is used to boost the tx power of lower
rates. Take this into account when parsing rate power.
The assumption here is, that the first rate (OFDM 6M or CCK 1M) has the
highest tx power
Signed-off-by: Felix Fietkau <[email protected]>
A switch from one rate index to the next only happens when tx count from
the current slot is greater than MT7615_RATE_RETRY, which is 1 has to be
subtracted from count, instead of added to it.
Signed-off-by: Felix Fietkau <[email protected]>
A switch from one rate index to the next only happens when tx count from
the current slot is greater than MT7615_RATE_RETRY, which is 1 has to be
subtracted from count, instead of added to it.
Signed-off-by: Felix Fietkau <[email protected]>
Sometimes mt7615_mac_set_timing gets called while the slot time is still
configured to 20. Ensure that in this case it always uses the OFDM CFend
rate.
Signed-off-by: Felix Fietkau <[email protected]>
Sometimes mt7615_mac_set_timing gets called while the slot time is still
configured to 20. Ensure that in this case it always uses the OFDM CFend
rate.
Signed-off-by: Felix Fietkau <[email protected]>
In VHT mode we can receive larger MPDUs. Increasing the buffer size reduces
fragmentation here, which should improve performance.
Signed-off-by: Felix Fietkau <[email protected]>
The hardware requirements for A-MPDU density are entirely on the tx side,
not the rx side. Because of that, the IE value should stay at 0 and the
minimum value should instead be enforced in WTBL/TXWI
MT7615 has no restrictions here
Signed-off-by: Felix Fietkau <[email protected]>
With MSI interrupts, IRQs must not be enabled from within the IRQ handler,
because that can lead to lost events.
Defer IRQ processing to a tasklet, which is also responsible for enabling
IRQs (to avoid race conditions against the handler)
Co-developed-by: Soul Huang <[email protected]>
Signed-off-by: Felix Fietkau <[email protected]>
Avoid having to attempt hw scan and fall back to software for every scan
on devices/firmware without hw scan support
Signed-off-by: Felix Fietkau <[email protected]>
The reference driver does not seem to enable it by default, only under certain
conditions, e.g. when a .bin file is loaded.
Make it opt-in via a device tree property for now, in case it is needed on some
boards.
Signed-off-by: Felix Fietkau <[email protected]>
The reference driver does not seem to enable it by default, only under certain
conditions, e.g. when a .bin file is loaded.
Make it opt-in via a device tree property for now, in case it is needed on some
boards.
Signed-off-by: Felix Fietkau <[email protected]>
When the EEPROM data is read from flash, it can contain tx DPD calibration
data. Add support for sending the data to the firmware.
Signed-off-by: Felix Fietkau <[email protected]>
When the EEPROM data is read from flash, it can contain DC offset calibration
data. Add support for sending the data to the firmware.
Signed-off-by: Felix Fietkau <[email protected]>
The reference driver does not seem to enable it by default, only under certain
conditions, e.g. when a .bin file is loaded.
Make it opt-in via a device tree property for now, in case it is needed on some
boards.
Signed-off-by: Felix Fietkau <[email protected]>
Felix Fietkau