netdev - Re: [PATCH net-next v3 1/4] net: stmmac: Switch to zero-copy in non-XDP RX path

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <Z5X1M0Fs-K6FkSAl@shredder>
Date: Sun, 26 Jan 2025 10:41:23 +0200
From: Ido Schimmel <idosch@...sch.org>
To: Furong Xu <0x1207@...il.com>
Cc: Andrew Lunn <andrew@...n.ch>, Brad Griffis <bgriffis@...dia.com>,
	Jon Hunter <jonathanh@...dia.com>, netdev@...r.kernel.org,
	linux-stm32@...md-mailman.stormreply.com,
	linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
	Alexander Lobakin <aleksander.lobakin@...el.com>,
	Joe Damato <jdamato@...tly.com>,
	Andrew Lunn <andrew+netdev@...n.ch>,
	"David S. Miller" <davem@...emloft.net>,
	Eric Dumazet <edumazet@...gle.com>,
	Jakub Kicinski <kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>,
	Maxime Coquelin <mcoquelin.stm32@...il.com>, xfr@...look.com,
	"linux-tegra@...r.kernel.org" <linux-tegra@...r.kernel.org>
Subject: Re: [PATCH net-next v3 1/4] net: stmmac: Switch to zero-copy in
 non-XDP RX path

Hi,

On Sat, Jan 25, 2025 at 10:43:42PM +0800, Furong Xu wrote:
> Hi Ido
> 
> On Sat, 25 Jan 2025 12:20:38 +0200, Ido Schimmel wrote:
> 
> > On Fri, Jan 24, 2025 at 10:42:56AM +0800, Furong Xu wrote:
> > > On Thu, 23 Jan 2025 22:48:42 +0100, Andrew Lunn <andrew@...n.ch>
> > > wrote: 
> > > > > Just to clarify, the patch that you had us try was not intended
> > > > > as an actual fix, correct? It was only for diagnostic purposes,
> > > > > i.e. to see if there is some kind of cache coherence issue,
> > > > > which seems to be the case?  So perhaps the only fix needed is
> > > > > to add dma-coherent to our device tree?    
> > > > 
> > > > That sounds quite error prone. How many other DT blobs are
> > > > missing the property? If the memory should be coherent, i would
> > > > expect the driver to allocate coherent memory. Or the driver
> > > > needs to handle non-coherent memory and add the necessary
> > > > flush/invalidates etc.  
> > > 
> > > stmmac driver does the necessary cache flush/invalidates to
> > > maintain cache lines explicitly.  
> > 
> > Given the problem happens when the kernel performs syncing, is it
> > possible that there is a problem with how the syncing is performed?
> > 
> > I am not familiar with this driver, but it seems to allocate multiple
> > buffers per packet when split header is enabled and these buffers are
> > allocated from the same page pool (see stmmac_init_rx_buffers()).
> > Despite that, the driver is creating the page pool with a non-zero
> > offset (see __alloc_dma_rx_desc_resources()) to avoid syncing the
> > headroom, which is only present in the head buffer.
> > 
> > I asked Thierry to test the following patch [1] and initial testing
> > seems OK. He also confirmed that "SPH feature enabled" shows up in the
> > kernel log.
> > BTW, the commit that added split header support (67afd6d1cfdf0) says
> > that it "reduces CPU usage because without the feature all the entire
> > packet is memcpy'ed, while that with the feature only the header is".
> > This is no longer correct after your patch, so is there still value in
> > the split header feature? With two large buffers being allocated from
> 
> Thanks for these great insights!
> 
> Yes, when "SPH feature enabled", it is not correct after my patch,
> pp_params.offset should be updated to match the offset of split payload.
> 
> But I would like to let pp_params.max_len remains to
> dma_conf->dma_buf_sz since the sizes of both header and payload are
> limited to dma_conf->dma_buf_sz by DMA engine, no more than
> dma_conf->dma_buf_sz bytes will be written into a page buffer.
> So my patch would be like [2]:
> 
> BTW, the split header feature will be very useful on some certain
> cases, stmmac driver should support this feature always.
> 
> [2]
> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> index edbf8994455d..def0d893efbb 100644
> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> @@ -2091,7 +2091,7 @@ static int __alloc_dma_rx_desc_resources(struct stmmac_priv *priv,
>         pp_params.nid = dev_to_node(priv->device);
>         pp_params.dev = priv->device;
>         pp_params.dma_dir = xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
> -       pp_params.offset = stmmac_rx_offset(priv);
> +       pp_params.offset = priv->sph ? 0 : stmmac_rx_offset(priv);

SPH is the only scenario in which the driver uses multiple buffers per
packet?

>         pp_params.max_len = dma_conf->dma_buf_sz;

Are you sure this is correct? Page pool documentation says that "For
pages recycled on the XDP xmit and skb paths the page pool will use the
max_len member of struct page_pool_params to decide how much of the page
needs to be synced (starting at offset)" [1].

While "no more than dma_conf->dma_buf_sz bytes will be written into a
page buffer", for the head buffer they will be written starting at a
non-zero offset unlike buffers used for the data, no?

[1] https://docs.kernel.org/networking/page_pool.html#dma-sync