Re: [PATCH wireless-drivers 2/2] mt76: dma: fix buffer unmap with non-linear skbs

[Lorenzo Bianconi <lorenzo@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 2359 bytes --]

On Oct 24, Felix Fietkau wrote:
> On 2019-10-24 00:23, Lorenzo Bianconi wrote:
> > mt76 dma layer is supposed to unmap skb data buffers while keep txwi
> > mapped on hw dma ring. At the moment mt76 wrongly unmap txwi or does
> > not unmap data fragments in even positions for non-linear skbs. This
> > issue may result in hw hangs with A-MSDU if the system relies on IOMMU
> > or SWIOTLB. Fix this behaviour properly unmapping data fragments on
> > non-linear skbs.
> > 
> > Fixes: 17f1de56df05 ("mt76: add common code shared between multiple chipsets")
> > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > ---
> >  drivers/net/wireless/mediatek/mt76/dma.c | 10 +++++++---
> >  1 file changed, 7 insertions(+), 3 deletions(-)
> > 
> > diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c
> > index c747eb24581c..8c27956875e7 100644
> > --- a/drivers/net/wireless/mediatek/mt76/dma.c
> > +++ b/drivers/net/wireless/mediatek/mt76/dma.c
> > @@ -93,11 +93,14 @@ static void
> >  mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
> >  			struct mt76_queue_entry *prev_e)
> >  {
> > -	struct mt76_queue_entry *e = &q->entry[idx];
> >  	__le32 __ctrl = READ_ONCE(q->desc[idx].ctrl);
> > +	struct mt76_queue_entry *e = &q->entry[idx];
> >  	u32 ctrl = le32_to_cpu(__ctrl);
> > +	bool mcu = e->skb && !e->txwi;
> > +	bool first = e->skb == DMA_DUMMY_DATA || e->txwi == DMA_DUMMY_DATA ||
> > +		     (e->skb && !skb_is_nonlinear(e->skb));
> It seems to me that these conditions could be true not just for the
> first entry, but also following entries except for the last one.
> I think we should add a 'bool has_txwi' field in struct mt76_queue_entry
> to indicate that the first dma address points to a txwi that should not
> be unmapped.

I agree 'first' is misleading since this condition is used to unamp even the
very last data fragment if we have an even number of data fragments
(e.g linear area + one fragment). For the following entries except for the last
one 'first' is false since e->skb is NULL (e->skb is not NULL just for the very
last entry), right? Btw we can remove mcu bool.
In order to improve code readability I agree to add a bool in mt76_queue_entry to
unmap buf0. I will fix it in v2.

Regards,
Lorenzo

> 
> - Felix

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]