When receiving an A-MSDU containing only a single subframe, status->last_amsdu
is set, but no previous head is present in the rx A-MSDU queue.
In this case, the A-MSDU subframe will be held until the next one is received,
potentially causing significant extra latency.
Rework the code to make it easier to read and less convoluted, and release the
newly created head if status->last_amsdu is set.
Signed-off-by: Felix Fietkau <[email protected]>
After chip reset, the DMA scheduler needs to be initialized as well.
Since the code is PCI/SoC specific, move it to pci_mac.c, so that it
can depend on a function in dma.c
Signed-off-by: Felix Fietkau <[email protected]>
When disassociating and clearing the BSSID of a repeater entry used by a client
mode interface, the corresponding MAC address entry can get lost too, causing
the client interface to not be able to associate anymore.
Fix this by also refreshing the MAC address when setting the BSSID
Signed-off-by: Felix Fietkau <[email protected]>
With buf uninitialized in mt76_dma_tx_queue_skb_raw, its field skip_unmap
could potentially inherit a non-zero value from stack garbage.
If this happens, it will cause DMA mappings for MCU command frames to not be
unmapped after completion
Fixes: 27d5c528a7ca ("mt76: fix double DMA unmap of the first buffer on 7615/7915")
Cc: [email protected]
Signed-off-by: Felix Fietkau <[email protected]>
Deleting a key with the previous key index deletes the current key
Rework the code to better keep track of multiple keys and check for the
key index before deleting the current key
Signed-off-by: Felix Fietkau <[email protected]>
When running out of room in the tx queue after calling drv->tx_prepare_skb,
the buffer list will already have been modified on MT7615 and newer drivers.
This can leak a DMA mapping and will show up as swiotlb allocation failures
on x86.
Fix this by moving the queue length check further up. This is less accurate,
since it can overestimate the needed room in the queue on MT7615 and newer,
but the difference is small enough to not matter in practice.
Signed-off-by: Felix Fietkau <[email protected]>
For AP and Client mode, the hardware can pass received packets as 802.3 frames
that can be passed to the network stack as-is.
Signed-off-by: Felix Fietkau <[email protected]>
The first pointer in the txp needs to be unmapped as well, otherwise it will
leak DMA mapping entries
Fixes: 27d5c528a7ca ("mt76: fix double DMA unmap of the first buffer on 7615/7915")
Signed-off-by: Felix Fietkau <[email protected]>
The first pointer in the txp needs to be unmapped as well, otherwise it will
leak DMA mapping entries
Fixes: 27d5c528a7ca ("mt76: fix double DMA unmap of the first buffer on 7615/7915")
Signed-off-by: Felix Fietkau <[email protected]>
There were multiple issues in the current key set/remove code:
- deleting a key with the previous key index deletes the current key
- BIP key would only be uploaded correctly initially and corrupted on rekey
Rework the code to better keep track of multiple keys and check for the
key index before deleting the current key
Signed-off-by: Felix Fietkau <[email protected]>
With a maximum length of 7991, the radio sometimes locks up under load when
transmitting A-MSDU frames to lots of stations. Using the lower limit makes
it work reliably again
Signed-off-by: Felix Fietkau <[email protected]>
Instead of holding it for the duration of an entire station schedule run,
which can block out competing tasks for a significant amount of time,
only hold it for scheduling one batch of packets for one station.
Improves responsiveness under load
Signed-off-by: Felix Fietkau <[email protected]>
For performance reasons, mt7915 supports using 2 PCIE gen1 links on
platforms that don't support gen2.
Add support for using this to move traffic for a second DBDC band onto
a dedicated link
Signed-off-by: Felix Fietkau <[email protected]>
Since vif indexes are per-phy, we need to factor in the phy index to avoid
collisions on the WTBL index
Signed-off-by: Felix Fietkau <[email protected]>
Only clear dev->phy2 after the phy is gone, the driver may still need to access
it until shutdown is complete
Signed-off-by: Felix Fietkau <[email protected]>
When the EEPROM band fields contain default values, assign 2.4 GHz to the
first band and 5 GHz to the second.
Signed-off-by: Felix Fietkau <[email protected]>
It causes high CPU load on the WA core, which can lead to extra latency when
using many stations. Dropping packets for long queues is managed by mac80211
fq_codel instead
Signed-off-by: Felix Fietkau <[email protected]>
Add support for passing flags for selecting the MCU target and query type
instead of trying to detect it based on the command id
Signed-off-by: Felix Fietkau <[email protected]>
This is causing some extreme latency spikes when sending traffic to
multiple stations at the same time. Disable this until the firmware issue
is sorted out
Signed-off-by: Felix Fietkau <[email protected]>
Instead of stopping and waking only a single queue, handle all phy tx queues
mapped ot the same hardware queue.
Also allow the driver to block tx queues
Signed-off-by: Felix Fietkau <[email protected]>
It was accidentally dropped while adding multiple wiphy support
Fixes fast-rx support and avoids handling reordering in both mac80211
and the driver
Cc: [email protected]
Fixes: c89d36254155 ("mt76: add function for allocating an extra wiphy")
Signed-off-by: Felix Fietkau <[email protected]>
When per-packet rate reporting is enabled, the hardware can get stuck under
some conditions. It self-recovers quickly, but in practice this leads to
reduced performance.
In order to avoid running into this issue, disable per-packet rate reporting
by default and query the rx rate from firmware instead, unless monitor mode
is enabled.
Signed-off-by: Felix Fietkau <[email protected]>
Can be used by the caller to get the response data directly instead of using the
hack of storing the result in internal data structures from .mcu_parse_response
Signed-off-by: Felix Fietkau <[email protected]>
The firmware/hardware supports only 16 entries. Trying to use more can lead
to timeouts on some firmware versions
Signed-off-by: Felix Fietkau <[email protected]>
In the MT7915 info, the fields for the WLAN index / queue overlap with the
token id, and the MT_TX_FREE_PAIR bit indicates, which one is present.
If MT_TX_FREE_PAIR is set, skip processing the token index, since the data
will not be valid.
This fixes accidentally freeing tokens which are still in use by the
hardware with a wrong station pointer.
Signed-off-by: Felix Fietkau <[email protected]>
When looking for a MAC address slot, start by using main BSSID slots 1-3,
afterwards use 16 repeater mode BSSID slots, then start using the slots
usually used for AP mode.
This search order should prevent unnecessary conflicts with AP mode interfaces
on the same PHY
Signed-off-by: Felix Fietkau <[email protected]>
The newer chipsets can deal with the same STA on multiple interfaces
Preparation for supporting more station interfaces
Signed-off-by: Felix Fietkau <[email protected]>
This is primarily for testing and can be used in combination with monitor
mode to make the card respond to packets sent to a specific MAC address.
For now this is only exposed as a debug/testing feature, later on the
approach might be used to support more concurrent station interfaces
Signed-off-by: Felix Fietkau <[email protected]>
After fixing mac80211 to allow larger A-MSDUs in some cases, there have been
reports of performance regressions and packet loss with some clients.
It appears that the issue occurs when the hardware is transmitting A-MSDUs
bigger than 8k. Limit the local VHT MPDU size capability to 7991, matching
the value used for MT7915 as well.
Signed-off-by: Felix Fietkau <[email protected]>
On MT7615 and MT7915, packets received from DMA can include TX_FREE events,
which free up space for tx. In order to process those packets faster, keep
keep a longer internal queue of packets and pass as many packets to the
stack as the NAPI budget allows
Signed-off-by: Felix Fietkau <[email protected]>
Since the switch to using AQL by default, mtxq->retry_q is never filled anymore
Remove it to get rid of a few more unnecessary cycles in the tx path
Signed-off-by: Felix Fietkau <[email protected]>
Compared to mac80211 ACs, MT7915 queue numbers are in reverse order
There is no need for the defensive WARN_ON_ONCE, so we can simplify
the function to avoid the array lookup
Signed-off-by: Felix Fietkau <[email protected]>
Do not alter the tx/rx chain settings during channel setup, antennas are
remapped by the testmode specific register writes already
Signed-off-by: Felix Fietkau <[email protected]>
Some devices using MT7628 or MT7603 have only one antenna chain connected.
Detect these using the EEPROM rx/tx path settings
Reported-by: Qin Wei <[email protected]>
Signed-off-by: Felix Fietkau <[email protected]>
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]>
Felix Fietkau