Linux only sets that field for non-ht, and leaves the field blank for ht. Linux expects lower level set it. This is realized after checking the mac80211 source code. After studying the duration/id field of COTS WiFi chip by wireshark, we calculate it in driver sdr.c gen_ht_duration_id(). The concrete logic/method of the calculation is in the comments of the function gen_ht_duration_id().
Don't need to reset _prev variables every time when it is not ht aggr qos data. Reason:
1. In 99.9999% cases, the ht always use qos data and goes to prio/queue_idx 2. By not resetting the variable to -1, we can have continuous aggregation packet operation in FPGA queue 2.
2. In other words, the aggregation operation for queue 2 in FPGA won't be interrupted by other non aggregation packets (control/management/beacon/etc.) that go to queue 0 (or other queues than 2).
3. From wired domain and upper level ( DSCP, AC (0~3), WMM management, 802.11D service classes and user priority (UP) ) to chip/FPGA queue index, thre should be some (complicated) mapping relationship.
4. More decent design is setting these aggregation flags (ht_aggr_start) per queue/prio here in driver. But since now only queue 2 and 0 are used (data goes to queue 2, others go to queue 0) in normal (most) cases, let's not go to the decent (complicated) solution immediately.
- Manually issue Tx Quadrature calibration if frequency change is more than 100MHz
- Disable FPGA SPI module before calibration
- Add xpu reg 13 to disable control manually
- 2022-03-04 detailed test result:
- - bb_gain 290 work for 11a/g all mcs
- - bb_gain 290 work for 11n mcs 1~7 (aggr and non aggr)
- - bb_gain 290 destroy 11n mcs 0 long (MTU 1500) tx pkt due to high PAPR (Peak to Average Power Ratio)
- - bb_gain 250 work for 11n mcs 0
So, a conservative bb_gain 250 should be used
1. mixer/duc is not needed because we will not use offset tuning after the ad9361 tx lo control via FPGA is supported.
2. source selection register is not needed as well.
3. arbitrary IQ register is added.
Some client, like iPhone, always has frequent PS (Power Saving) state change like this:
sdr0: STA e2:72:49:82:a6:a0 aid 1 enters power save mode
sdr0: STA e2:72:49:82:a6:a0 aid 1 exits power save mode
sdr0: STA e2:72:49:82:a6:a0 aid 1 sending 0 filtered/0 PS frames since STA woke up
Now with these correct hw flag setting, the link is more stable
1. test_mode 0 -- normal, no aggregation (AMPDU); test_mode 1 -- experimental, has AMPDU support
2. init_tx_att -- initial tx attenuation in dB*1000 format. example: put 20000 for 20dB
By default the TID is disabled in FPGA, because we currently try to TX and RX traffic for all TIDs. So, the TID related operations in sdr.c are removed.
1. Remove the last_auto_fpga_lbt_th saving from sdrctl set reg command. Otherwise, repeated sdrctl set reg will save wrong value into last_auto_fpga_lbt_th
2. The last_auto_fpga_lbt_th is only set in ad9361_rf_set_channel, which is called at least once by Linux after NIC is up
1. The issue only happens at zcu102 side, when it is tested as AP together with zedboard
2. The issue does not happen when zcu102 is client and zedboard is AP
3. The issue (most likely) does not happen in places other than iGent (like Pablo home)
4. Sometimes it does happen at my home when I test zcu102 as AP together with COTS WiFi
5. Indeed seems like the environment related. Guess some quick small packets in the environment quickly flush/round-up/mess-up the rx dma cyclic buffer, and the rx interrupt internal static variable target_buf_idx_old loses track of the background automatic rx dma cyclic buffer
6. The fix is for all board types (zcu102, zedboard, 7035, etc)
7. The driver compiling make_all.sh script generates USE_NEW_RX_INTERRUPT macro to pre_def.h to enable the new code (while keeping the old code). You can use the script as before.
8. The logic of the fix is that exhaustive search all the rx dma cyclic buffer in rx interrupt to get packet to Linux in the first place.
we do not maintain our own (modified) ad9361 driver anymore! The original ad9361 driver in the Linux kernel tree can be used with some extra EXPORT_SYMBOL.
