On Wed, Jun 12, 2019 at 04:44:01PM +0200, Lorenzo Bianconi wrote: > On Jun 12, Stanislaw Gruszka wrote: > > On Wed, Jun 12, 2019 at 02:59:05PM +0200, Stanislaw Gruszka wrote: > > > > > If max RX AMSDU size is 3839B I do not see reason why we allocate > > > > > MT_SG_MAX_SIZE=8 of MT_RX_BUF_SIZE=2k buffers for sg_en case. > > > > > I thought the reason is that max AMSDU size is 16kB so it fit into > > > > > 8 sg buffers of 2k. > > > > > > > > > > In other words, for me, looks like either > > > > > - we can not handle AMSDU for non sg case because we do not > > > > > allocate big enough buffer > > > > > > > > I think AMSDU is mandatory and we currently support it even for non-sg case > > > > (since max rx AMSDU is 3839B) > > > > > > > > > or > > > > > - we can just use one PAGE_SIZE buffer for rx and remove sg > > > > > buffers for rx completely > > > > > > > > using sg buffers we can support bigger rx AMSDU size in the future without using > > > > huge buffers (e.g. we can try to use IEEE80211_MAX_MPDU_LEN_HT_7935 with > > > > mt76x2u) > > > > > > I think it would be simpler just to allocate 2 pages for 7935B . > > > > And if we could determine that there is no true need to use sg for rx, > > I think best approach here would be revert f8f527b16db5 in v5.2 to fix > > regression and remove rx sg in -next. That would make code simpler, > > allocate 4k instead 16k per packet, allow to use build_skb (4096 - 3839 > > give enough space for shared info) and not use usb hcd bounce buffer. > > I do not think we should drop sg support since: > - it allow us to rx huge amsdu frames (e.g. IEEE80211_MAX_MPDU_LEN_VHT_11454) > using multiple one page buffer. I think there will be new usb devices where we can > increase amsdu size (we can increase it even on mt76x2u usb 3.0 devices) I looked at intel wifi drivers and this is handled by amsdu_size module parameter, supported values are 4k, 8k and 12k. RX allocation size and proper values in vht_cap & ht_cap are set accordingly. Assuming (some) mt76 HW and FW can handle bigger AMSDUs I think we should do similar thing. Otherwise looks for me, we just waste memory and have not needed code for no true reason. > space needed for skb_shared_info is 320B on a x86_64 device Uhh, I haven't expected that sk_shared_info() is that big, so indeed build_skb could not used and 128B copy fallback will be necessary. Stanislaw
