[0/6] net: migrate NO_NET out of the networking stack choice

[Quentin Schulz]

Hi Simon,

On 4/20/26 9:51 PM, Simon Glass wrote:
> Hi Quentin,
> 
> On 2026-04-20T11:36:06, Quentin Schulz <foss+uboot at 0leil.net> wrote:
> 
>> This series introduces a new NET umbrella symbol with NET_LEGACY and NET_LWIP underneath.
> 
> Thanks for this cleanup!
> 
> Re SYS_RX_ETH_BUFFER, how about moving the extern for net_rx_packets[]
> to net-legacy.h ? (is it used with lwip?)
> 

That's not the only thing making use of SYS_RX_ETH_BUFFER.

- drivers/net/fsl_enetc.h uses it as value for ENETC_BD_CNT which is 
used in drivers/net/fsl_enetc.c,
- drivers/net/rtl8169.c uses it (if defined, otherwise defaults to 4) to 
create a static buffer, see NUM_RX_DESC,
- include/net-common.h sets PKTBUFSRX to it, this in turn is used in 
*many* places, including many network (assumed Ethernet controller) 
drivers, net/lwip/net-lwip.c, net/net-common.c (with net_rx_packets 
array, used all over the place), net/net.c, and net/tcp.c.

So seems to be used for both NET_LEGACY and NET_LWIP.

Simply guarding the "#include <net-common.h>" with 
CONFIG_IS_ENABLED(NET) won't do as many files currently expect some 
functions/macros/constants from that header to be defined.

Another option is to ifdef PKTBUFSRX contant and net_rx_packets array in 
include/net-common.h. But it's used in arch/sandbox/include/asm/eth.h 
and drivers/dma/ti/k3-udma.c which aren't clearly dpeending on NET being 
set. All defconfigs enabling CONFIG_TI_K3_NAVSS_UDMA do not modify the 
default CONFIG_SYS_RX_ETH_BUFFER which is 4 (according to 
./tools/qconfig.py -l -f CONFIG_TI_K3_NAVSS_UDMA 
'~CONFIG_SYS_RX_ETH_BUFFER=4') so it should be fine for this driver not 
to have PKTBUFSRX defined (as UDMA_RX_DESC_NUM will then default to 4). 
According to ./tools/qconfig.py -l -f CONFIG_SANDBOX CONFIG_NO_NET, all 
sandbox defconfigs have network enabled so maybe 
arch/sandbox/include/asm/eth.h is not an issue either (or rather, for 
later :) ).

With

diff --git a/include/net-common.h b/include/net-common.h
index 69b6316c1ec..0c260873c2c 100644
--- a/include/net-common.h
+++ b/include/net-common.h
@@ -20,7 +20,9 @@
   *	alignment in memory.
   *
   */
+#if CONFIG_IS_ENABLED(NET)
  #define PKTBUFSRX	CONFIG_SYS_RX_ETH_BUFFER
+#endif
  #define PKTALIGN	ARCH_DMA_MINALIGN

  /* IPv4 addresses are always 32 bits in size */
@@ -132,7 +134,9 @@ static inline void net_set_state(enum net_loop_state 
state)
  }

  extern int		net_restart_wrap;	/* Tried all network devices */
+#if CONFIG_IS_ENABLED(NET)
  extern uchar		*net_rx_packets[PKTBUFSRX]; /* Receive packets */
+#endif
  extern const u8		net_bcast_ethaddr[ARP_HLEN];	/* Ethernet broadcast 
address */
  extern struct in_addr	net_ip;		/* Our    IP addr (0 = unknown) */
  /* Indicates whether the pxe path prefix / config file was specified 
in dhcp option */
diff --git a/net/Kconfig b/net/Kconfig
index e712a0dd2ac..deab340f26b 100644
--- a/net/Kconfig
+++ b/net/Kconfig
@@ -279,8 +279,6 @@ config TFTP_BLOCKSIZE
  	  almost-MTU block sizes.
  	  You can also activate CONFIG_IP_DEFRAG to set a larger block.

-endif   # if NET
-
  config SYS_RX_ETH_BUFFER
          int "Number of receive packet buffers"
          default 4
@@ -289,3 +287,6 @@ config SYS_RX_ETH_BUFFER
            controllers it is recommended to set this value to 8 or even 
higher,
            since all buffers can be full shortly after enabling the 
interface on
            high Ethernet traffic.
+
+endif   # if NET
+


And building sandbox_defconfig with CONFIG_NO_NET=y, the build finishes 
without any issue. So maybe that's the way forward? I think this would 
be better as a separate series though, what do you think?

Cheers,
Quentin