Re: [PATCH net v3 5/6] net: stmmac: size the RX buffers from the frame length, not the MTU

[Paolo Abeni <pabeni@redhat.com>]

This is an AI-generated review of your patch. The human sending this
email has considered the AI review valid, or at least plausible.
Full review at: https://netdev-ai.bots.linux.dev/sashiko/
---
net: stmmac: size the RX buffers from the frame length, not the MTU

The patch stops feeding the bare MTU into the RX buffer size selection and
instead computes a frame length that includes the L2 header, two VLAN tags,
the FCS and NET_IP_ALIGN.  stmmac_set_bfsize() is renamed to take a length
and its boundary comparisons are switched from '>=' to '>'.

A few questions below.

> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> index 62c3441911e71..29b970574a380 100644
> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> @@ -1536,17 +1536,17 @@ static unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
>  	return NET_SKB_PAD + NET_IP_ALIGN;
>  }
>  
> -static int stmmac_set_bfsize(int mtu)
> +static int stmmac_set_bfsize(int len)
>  {
>  	int ret;
>  
> -	if (mtu >= BUF_SIZE_8KiB)
> +	if (len > BUF_SIZE_8KiB)
>  		ret = BUF_SIZE_16KiB;

[Severity: High]
This isn't a bug introduced by this patch, but the 16 KiB selection here is
reachable in configurations whose descriptors cannot express a buffer that
large, and this patch moves the threshold about 28 bytes of MTU lower.

In chain mode mac->mode points at chain_mode_ops, which does not implement
set_16kib_bfsize, so stmmac_do_callback() returns -EINVAL, bfsize is forced
to 0 and stmmac_set_bfsize(len) runs unconditionally.  For enhanced
descriptors the programmed buffer 1 is clamped:

drivers/net/ethernet/stmicro/stmmac/enh_desc.c:enh_desc_init_rx_desc() {
	bfsize1 = min(bfsize, BUF_SIZE_8KiB);
	p->des1 |= cpu_to_le32(bfsize1 & ERDES1_BUFFER1_SIZE_MASK);
}

and for normal descriptors:

drivers/net/ethernet/stmicro/stmmac/norm_desc.c:ndesc_init_rx_desc() {
	bfsize1 = min(bfsize, BUF_SIZE_2KiB - 1);
	p->des1 |= cpu_to_le32(bfsize1 & RDES1_BUFFER1_SIZE_MASK);
}

while the RX path still reports the software value for every non-last
descriptor:

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c:stmmac_rx_buf1_len() {
	/* First descriptor, not last descriptor and not split header */
	if (status & rx_not_ls)
		return priv->dma_conf.dma_buf_sz;
}

That value then drives dma_sync_single_range_for_cpu() and
skb_add_rx_frag().  Can this hand the stack more bytes than the engine
wrote, padding the frame with recycled page pool contents?  Chain mode is
reachable through the chain_mode module parameter and is forced for
dwmac-sun8i, which uses normal descriptors.

> -	else if (mtu >= BUF_SIZE_4KiB)
> +	else if (len > BUF_SIZE_4KiB)
>  		ret = BUF_SIZE_8KiB;
> -	else if (mtu >= BUF_SIZE_2KiB)
> +	else if (len > BUF_SIZE_2KiB)
>  		ret = BUF_SIZE_4KiB;
> -	else if (mtu > DEFAULT_BUFSIZE)
> +	else if (len > DEFAULT_BUFSIZE)
>  		ret = BUF_SIZE_2KiB;

[Severity: High]
This is a pre-existing issue and not something this patch introduces, but
the new thresholds widen the MTU range in which it applies (the step up to
BUF_SIZE_2KiB/BUF_SIZE_4KiB now happens roughly 28 bytes of MTU earlier).

For normal descriptors in ring mode, a second data buffer is advertised
whose bus address is never programmed:

drivers/net/ethernet/stmicro/stmmac/descs_com.h:ndesc_rx_set_on_ring() {
	if (bfsize >= BUF_SIZE_2KiB) {
		int bfsize2;

		bfsize2 = min(bfsize - BUF_SIZE_2KiB + 1, BUF_SIZE_2KiB - 1);
		p->des1 |= cpu_to_le32(FIELD_PREP(RDES1_BUFFER2_SIZE_MASK,
						  bfsize2));
	}
}

RDES3 is only ever written through ring_mode.c init_desc3()/refill_desc3(),
and the driver only calls those when dma_buf_sz equals BUF_SIZE_16KiB:

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c:stmmac_init_rx_buffers() {
	if (dma_conf->dma_buf_sz == BUF_SIZE_16KiB)
		stmmac_init_desc3(priv, p);
}

Devices on ndesc_ops are capped at max_mtu = SKB_MAX_HEAD(NET_SKB_PAD +
NET_IP_ALIGN), so they can only pick 2048 or 4096 and RDES3 stays 0 from
dma_alloc_coherent().  Since the second-address-chained bit is not set in
ring mode, does the RX DMA spill the part of the frame beyond the 2047-byte
buffer 1 to bus address 0?  This is driven by an inbound frame on cores
using normal descriptors (MAC100, or GMAC with plat->enh_desc == 0) with an
MTU above ~2 KiB.

>  	else
>  		ret = DEFAULT_BUFSIZE;
>  
> @@ -4063,7 +4063,7 @@ static struct stmmac_dma_conf *
>  stmmac_setup_dma_desc(struct stmmac_priv *priv, unsigned int mtu)
>  {
>  	struct stmmac_dma_conf *dma_conf;
> -	int bfsize, ret;
> +	int bfsize, len, ret;
>  	u8 chan;
>  
>  	dma_conf = kzalloc_obj(*dma_conf);
> @@ -4073,13 +4073,15 @@ stmmac_setup_dma_desc(struct stmmac_priv *priv, unsigned int mtu)
>  		return ERR_PTR(-ENOMEM);
>  	}
>  
> -	/* Returns 0 or BUF_SIZE_16KiB if mtu > 8KiB and dwmac4 or ring mode */
> -	bfsize = stmmac_set_16kib_bfsize(priv, mtu);
> +	len = mtu + ETH_HLEN + 2 * VLAN_HLEN + ETH_FCS_LEN + NET_IP_ALIGN;

[Severity: Low]
Should NET_IP_ALIGN be part of this sum?  It is head room that lives
outside the window described by dma_buf_sz, and it is already reserved
separately:

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c:__alloc_dma_rx_desc_resources() {
	dma_buf_sz_pad = stmmac_rx_offset(priv) + dma_conf->dma_buf_sz +
			 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
	...
	pp_params.offset = stmmac_rx_offset(priv);
	pp_params.max_len = dma_conf->dma_buf_sz;
}

with stmmac_rx_offset() returning NET_SKB_PAD + NET_IP_ALIGN, and
stmmac_init_rx_buffers() programming the descriptor at page dma address
plus buf->page_offset == stmmac_rx_offset(priv).

Counting it again inside len makes the selected buffer class, and therefore
the page pool order, depend on the architecture for a 2-byte wide MTU
window at each boundary.  For example MTU 4069/4070 with NET_IP_ALIGN == 2
gives len 4097/4098 and selects BUF_SIZE_8KiB, whereas the same MTU on
x86 (NET_IP_ALIGN == 0) selects BUF_SIZE_4KiB.

The commit message also lists "the NET_IP_ALIGN" alongside ETH_HLEN, the
VLAN tags and the FCS as if it were part of the frame.  Would it be more
accurate to describe it as head room, or drop the term?

> +
> +	/* Returns 0 or BUF_SIZE_16KiB if len > 8KiB and dwmac4 or ring mode */
> +	bfsize = stmmac_set_16kib_bfsize(priv, len);
>  	if (bfsize < 0)
>  		bfsize = 0;
>  
>  	if (bfsize < BUF_SIZE_16KiB)
> -		bfsize = stmmac_set_bfsize(mtu);
> +		bfsize = stmmac_set_bfsize(len);

[Severity: Medium]
The argument handed to the set_16kib_bfsize callback changes unit here from
MTU to frame length, but neither the ops prototype nor the two
implementations were updated.

hwif.h still declares:

	int (*set_16kib_bfsize)(int mtu);

dwmac4_descs.c:set_16kib_bfsize() keeps the inclusive test:

	if (unlikely(mtu >= BUF_SIZE_8KiB))
		ret = BUF_SIZE_16KiB;

while ring_mode.c:set_16kib_bfsize() uses the other operator:

	if (unlikely(mtu > BUF_SIZE_8KiB))
		ret = BUF_SIZE_16KiB;

Since dwmac4/dwmac5 use dwmac4_ring_mode_ops, does the new comment "Returns
0 or BUF_SIZE_16KiB if len > 8KiB" match what actually happens on those
cores?  A len of exactly BUF_SIZE_8KiB (MTU 8160 with NET_IP_ALIGN == 2)
fits an 8188-byte buffer and stmmac_set_bfsize() would return
BUF_SIZE_8KiB, but the unchanged '>=' returns BUF_SIZE_16KiB and the
'if (bfsize < BUF_SIZE_16KiB)' guard then skips the corrected helper, so
dma_buf_sz becomes 16368 instead of 8188.  That roughly doubles the pages
backing each RX buffer (order 2 to order 3) across the default
dma_rx_size of 512.

The commit message says the change lets us "use more consistent comparison
operators in stmmac_set_bfsize" - should the two set_16kib_bfsize
implementations and the 'mtu' parameter name be converted at the same time?
-- 
This is an AI-generated review.