On Fri, Jun 14, 2019 at 02:46:36PM +0200, Lorenzo Bianconi wrote:
> > >
> > > ack, right. I think patch 2/3 and 3/3 can go directly in Felix's tree
> > >
> > > >
> > > > > + int i, data_size;
> > > > >
> > > > > + data_size = rounddown(SKB_WITH_OVERHEAD(q->buf_size),
> > > > > + dev->usb.in_ep[MT_EP_IN_PKT_RX].max_packet);
> > > > > for (i = 0; i < nsgs; i++) {
> > > > > struct page *page;
> > > > > void *data;
> > > > > @@ -302,7 +304,7 @@ mt76u_fill_rx_sg(struct mt76_dev *dev, struct mt76_queue *q, struct urb *urb,
> > > > >
> > > > > page = virt_to_head_page(data);
> > > > > offset = data - page_address(page);
> > > > > - sg_set_page(&urb->sg[i], page, q->buf_size, offset);
> > > > > + sg_set_page(&urb->sg[i], page, data_size, offset);
> > > > <snip>
> > > > > - q->buf_size = dev->usb.sg_en ? MT_RX_BUF_SIZE : PAGE_SIZE;
> > > > > q->ndesc = MT_NUM_RX_ENTRIES;
> > > > > + q->buf_size = PAGE_SIZE;
> > > > > +
> > > >
> > > > This should be associated with decrease of MT_SG_MAX_SIZE to value that
> > > > is actually needed and currently this is 2 for 4k AMSDU.
> > >
> > > MT_SG_MAX_SIZE is used even on tx side and I do not think we will end up with a
> > > huge difference here
> >
> > So use different value as argument for mt76u_fill_rx_sg() in
> > mt76u_rx_urb_alloc(). After changing buf_size to PAGE_SIZE we will
> > allocate 8 pages per rx queue entry, but only 2 pages will be used
> > (with data_size change, 1 without data_size change). Or I'm wrong?
>
> yes, it is right (we will use two pages with data_size change). Maybe better to
> use 4 pages for each rx queue entry? (otherwise we will probably change it in
> the future)
We should not allocate more than is required. If support for bigger
rx AMSDUs will be added and announced in vht/ht capabilities to remote
stations, then increase of number of segments will be needed.
> > > > However I don't think allocating 2 pages to avoid ieee80211 header and SNAP
> > > > copy is worth to do. For me best approach would be allocate 1 page for
> > > > 4k AMSDU, 2 for 8k and 3 for 12k (still using sg, but without data_size
> > > > change to avoid 32B copying).
> > >
> > > From my point of view it is better to avoid copying if it is possible. Are you
> > > sure there is no difference?
> >
> > I do not understand what you mean by difference here.
>
> tpt differences, not sure if there are any
I would not expect any measurable difference in tpt nor in cpu usage
either way.
But I think, if some AMSDU subframe will be spited into two fragments,
data most likely will need to be linearised/copied, at some point before
passed to application, what will overcome any benefit of avoiding coping
802.11 header. Thought, I don't think this somehow will be visible in
benchmarking.
Stanislaw
