Re: git: 38cbdae33b7c - main - dwc: Rewrite barrier part

From: John Baldwin <jhb_at_FreeBSD.org>
Date: Tue, 10 Oct 2023 18:21:21 UTC
On 10/6/23 1:06 PM, Emmanuel Vadot wrote:
> The branch main has been updated by manu:
> 
> URL: https://cgit.FreeBSD.org/src/commit/?id=38cbdae33b7c3f772845c06f52b86c0ddeab6a17
> 
> commit 38cbdae33b7c3f772845c06f52b86c0ddeab6a17
> Author:     Emmanuel Vadot <manu@FreeBSD.org>
> AuthorDate: 2023-10-06 16:04:49 +0000
> Commit:     Emmanuel Vadot <manu@FreeBSD.org>
> CommitDate: 2023-10-06 17:05:45 +0000
> 
>      dwc: Rewrite barrier part
>      
>      We only need a barrier after writing the OWN bit so everything is
>      coherent for the DMA engine.
> ---
>   sys/dev/dwc/dwc1000_dma.c | 3 ---
>   1 file changed, 3 deletions(-)
> 
> diff --git a/sys/dev/dwc/dwc1000_dma.c b/sys/dev/dwc/dwc1000_dma.c
> index f6baa0bd3e45..4f8ae088305a 100644
> --- a/sys/dev/dwc/dwc1000_dma.c
> +++ b/sys/dev/dwc/dwc1000_dma.c
> @@ -216,7 +216,6 @@ txdesc_setup(struct dwc_softc *sc, int idx, bus_addr_t paddr,
>   	sc->txdesc_ring[idx].addr1 = (uint32_t)(paddr);
>   	sc->txdesc_ring[idx].desc0 = desc0;
>   	sc->txdesc_ring[idx].desc1 = desc1;
> -	wmb();
>   	sc->txdesc_ring[idx].desc0 |= TDESC0_OWN;
>   	wmb();
>   }
> @@ -237,7 +236,6 @@ rxdesc_setup(struct dwc_softc *sc, int idx, bus_addr_t paddr)
>   		sc->rxdesc_ring[idx].desc1 = ERDESC1_RCH |
>   		    MIN(MCLBYTES, ERDESC1_RBS1_MASK);
>   
> -	wmb();
>   	sc->rxdesc_ring[idx].desc0 = RDESC0_OWN;
>   	wmb();
>   	return (nidx);
> @@ -524,7 +522,6 @@ dma1000_rxfinish_locked(struct dwc_softc *sc)
>   
>   		m = dwc_rxfinish_one(sc, desc, sc->rxbuf_map + idx);
>   		if (m == NULL) {
> -			wmb();
>   			desc->desc0 = RDESC0_OWN;
>   			wmb();
>   		} else {

Hmm, are you sure?  I suspect you removed the wrong barrier here.  You want the
other writes to update other fields in the desc to post before the write that
sets TDESC0_OWN.  The barrier after TDESC0_OWN is useless, but the one you removed
is probably the one that matters.  In particular you don't want TDESC0_OWN to be
set when other writes (e.g. setting desc1) are not yet visible.

-- 
John Baldwin