On 19/03/2025 02:28, Ping-Ke Shih wrote: > Bitterblue Smith <rtl8821cerfe2@xxxxxxxxx> wrote: >> On 18/03/2025 04:06, Ping-Ke Shih wrote: >>> Bitterblue Smith <rtl8821cerfe2@xxxxxxxxx> wrote: >>>> On 17/03/2025 05:01, Ping-Ke Shih wrote: >>>>> Bitterblue Smith <rtl8821cerfe2@xxxxxxxxx> wrote: >>>>>> >>>>>> Tell the chip the maximum AMPDU size supported by the AP. This greatly >>>>>> improves the TX speed of RTL8814AU in the 2.4 GHz band. Before: ~90 >>>>>> Mbps. After: ~300 Mbps. >>>>>> >>>>>> Add this configuration for all the chips, even if it only has an effect >>>>>> on RTL8814AU in my tests. Surely they all need this. >>>>> >>>>> The hardware default value of REG_AMPDU_MAX_LENGTH is 0xffff (unlimited) >>>>> for most chips. It seems like RTL8812A/RTL8821A are also exceptions, so >>>>> at power on function they do >>>>> rtw_write32(rtwdev, REG_AMPDU_MAX_LENGTH, 0xffffffff); >>>>> >>>>> I feel RTL8814A has similar setting, so maybe you can just add similar >>>>> stuff. >>>>> >>>>> By the way, the AMPDU is controlled by TX descriptor basically: >>>>> pkt_info->ampdu_factor = ampdu_factor; >>>>> pkt_info->ampdu_density = ampdu_density; >>>>> pkt_info->ampdu_en = ampdu_en; >>>>> >>>>> Since you didn't change this part at all, I still feel setting >>>>> REG_AMPDU_MAX_LENGTH to 0xffffffff can fix low throughput problem. >>>>> >>>> >>>> I tried 0xffffffff just now and it doesn't work. It's the same with >>>> both of my routers. They advertise a maximum AMPDU size of 64 K. >>>> I can't just set it to 0xffff either, because then the upload speed >>>> in the 5 GHz band suffers a lot. The dual band router advertises a >>>> maximum AMPDU size of 256 K in the 5 GHz band so it gets a value of >>>> 0x3ffff. >>> >>> Not sure if 0xffffffff is a special value. Since this is a limit of >>> AMPDU length, you can set a constant large value such as 0x3ffff you >>> have tested. Is there special case it can't handle? >>> >>> >> >> 0x3ffff is not good for the 2.4 GHz band. The upload is only ~90 Mbps >> with both of the routers I tested. Same with 0x1ffff. Only 0xffff >> works well for them. > > Have you checked the packets in the air? How about their difference? > Intuitively larger REG_AMPDU_MAX_LENGTH would be better. > I checked today. With 0xffffffff I see 29% retransmission rate. With 0xffff the retransmission rate is only 8.6%. >> >> 0xffff is too little for the 5 GHz band. The upload speed is ~200 Mbps >> less than with 0x3ffff. >> >> I guess if you really don't want this patch I can hardcode 0xffff and >> 0x3ffff in rtw8814a_switch_band(). I just don't know if all access >> points will be happy with that. > > Initially I wanted to simply this patch, because changing REG_AMPDU_MAX_LENGTH > for other chips without testing is risky. With your tests, the behavior of > REG_AMPDU_MAX_LENGTH works not fully expected, so I suspect the risk > is even higher. > > Therefore, I would like limit this change to RTL8814A. Though hardcode proposal > is not sure workable for all AP, we also don't know this patch works for all > AP. Anyway this proposal is fine to me if we don't have other ideas. > > All right, I will limit this only to RTL8814A.
