On 2022-03-24 16:31, Christoph Hellwig wrote:
On Thu, Mar 24, 2022 at 05:29:12PM +0100, Maxime Bizon wrote:
I'm looking into this; but in the interest of a speedy resolution of
the regression I would be in favour of merging that partial revert
and reinstating it if/when we identify (and fix) any bugs in ath9k :)
This looks fishy:
ath9k/recv.c
/* We will now give hardware our shiny new allocated skb */
new_buf_addr = dma_map_single(sc->dev, requeue_skb->data,
common->rx_bufsize, dma_type);
if (unlikely(dma_mapping_error(sc->dev, new_buf_addr))) {
dev_kfree_skb_any(requeue_skb);
goto requeue_drop_frag;
}
/* Unmap the frame */
dma_unmap_single(sc->dev, bf->bf_buf_addr,
common->rx_bufsize, dma_type);
bf->bf_mpdu = requeue_skb;
bf->bf_buf_addr = new_buf_addr;
Creating a new mapping for the same buffer before unmapping the
previous one does looks rather bogus. But it does not fit the
pattern where revering the sync_single changes make the driver
work again.
OK, you made me look :)
Now that it's obvious what to look for, I can only conclude that during
the stanza in ath_edma_get_buffers(), the device is still writing to the
buffer while ownership has been transferred to the CPU, and whatever got
written while ath9k_hw_process_rxdesc_edma() was running then gets wiped
out by the subsequent sync_for_device, which currently resets the
SWIOTLB slot to the state that sync_for_cpu copied out. By the letter of
the DMA API that's not allowed, but on the other hand I'm not sure if we
even have a good idiom for "I can't tell if the device has finished with
this buffer or not unless I look at it" :/
Robin.
