> On Tue, Feb 12, 2019 at 03:25:30PM +0100, Lorenzo Bianconi wrote:
> > > On Tue, Feb 12, 2019 at 03:09:53PM +0100, Lorenzo Bianconi wrote:
> > > > >
> > > > > (repost with corrected Lorenzo email)
> > > > >
> > > > > On Tue, Feb 12, 2019 at 02:45:50PM +0100, Stanislaw Gruszka wrote:
> > > > > > On Tue, Feb 12, 2019 at 02:24:47PM +0100, LorenzoBianconilorenzo@xxxxxxxxxx wrote:
> > > > > > > From: Lorenzo Bianconi <lorenzo@xxxxxxxxxx>
> > > > > > >
> > > > > > > Use linear fragment and not a single usb scatter-gather buffer in mt76u
> > > > > > > {tx,rx} datapath if the usb controller has sg data length constraints.
> > > > > > > Moreover add disable_usb_sg module parameter in order to explicitly
> > > > > > > disable scatter-gather. SG I/O is not supported by all host drivers and
> > > > > > > some users have reported sg issues on AMD IOMMU.
> > > > > >
> > > > > > Again. This is not right approach. SG issues should be fixed
> > > > > > not workarounded.
> > > >
> > > > Hi Stanislaw,
> > > >
> > > > here we do not use SG, so num_sg is 0 and we use transfer_buffer. I do
> > > > not see how I am working around the issue.
> > >
> > > By avoiding SG buffer allocation and configuration which most likely
> > > need to be fixed.
> >
> > In my series I:
> > 1- set num_sg to 0
> > 2- use transfer_buffer
> >
> > please correct me if I am wrong but in your solution you did the same since AFAIK
> > PageHighMem is always 0 so you end up setting num_sg to 0 and using
> > transfer_buffer as well. Is my understanding correct?
>
> Yes. But it still using all existing SG allocation and setup code and
> buffer is tracked in urb->sg[0].
both of them are from page_frag_alloc()
>
> > > > Moreover with this approach we avoid some unnecessary operation in the hotpath
> > >
> > > What unnecessary operation ?
> >
> > the ones in mt76u_fill_bulk_urb()
>
> Your patches also add extra operations on hotpath due to urb->num_sgs and
> dev->usb.sg_en checks.
here I guess you should use mt76u_check_sg() instead of
udev->bus->sg_tablesize > 0 so I think you need to precompute it since it is in
the hotpath
Moreover the RFC series has been tested by multiple users and on multiple
devices
Regards,
Lorenzo
>
> Stanislaw
