> -----Original Message----- > From: Martin Blumenstingl <martin.blumenstingl@xxxxxxxxxxxxxx> > Sent: Tuesday, June 20, 2023 4:38 AM > To: Ping-Ke Shih <pkshih@xxxxxxxxxxx> > Cc: linux-wireless@xxxxxxxxxxxxxxx; Lukas F. Hartmann <lukas@xxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx; > tony0620emma@xxxxxxxxx; jernej.skrabec@xxxxxxxxx > Subject: Re: wifi: rtw88: question about SDIO RX aggregation limiting > > Hello Ping-Ke, > > apologies for the long delay. > > On Tue, Jun 13, 2023 at 4:20 AM Ping-Ke Shih <pkshih@xxxxxxxxxxx> wrote: > [...] > > The unit of BIT_RXDMA_AGG_PG_TH is 1k bytes, so I think you can > > set mmc_host->max_req_size/1024. > I tried this but I got a result that I don't understand. > I've been testing with three BIT_RXDMA_AGG_PG_TH values on a SoC that > can handle 255 * 1024 bytes. Each time I connected to the same AP and > downloaded a bigger file over http(s). > BIT_RXDMA_AGG_PG_TH: biggest observed rx_len in rtw_sdio_rxfifo_recv() > 255: 20968 > 6: 5122 > 1: 1602 Please also print out number of packets you receive, and then we can see how many packets aggregate. > > The biggest rx_len I have observed for BIT_RXDMA_AGG_PG_TH 1 looks suspicious: > My understanding is that I shouldn't be seeing rx_len larger than > BIT_RXDMA_AGG_PG_TH * 1024. > BIT_RXDMA_AGG_PG_TH = 6 is within this limit but BIT_RXDMA_AGG_PG_TH = > 1 isn't (I'm seeing 578 extra bytes in addition to the 1024 bytes that > I was expecting). Assume threshold is 1k, and single one packet is larger than 1k. Hardware will not split it into two. Also, please make sure 0x280[29] BIT_EN_PRE_CALC is 1. Otherwise, it will possibly aggregate additional one packet to over the threshold. 0x280[15:8] is timeout time in unit of 1us for SDIO interface. When set threshold to 255, you can enlarge this to see if it can aggregate more as expected. > Do you have any idea where this is coming from? I'm worried that we > can still end up with the problem that Lukas described but seems to > not have hit in his testing with BIT_RXDMA_AGG_PG_TH = 6 > > > I wonder why 0x6 works on Amlogic SoCs. Could you or Lukas compare performance > > between the settings of 0x1 and 0x6? > I can do this later this week but I'd like to understand the above > results first. > Ping-Ke
