Re: [PATCH net-next] net: lwtunnel: use GFP_KERNEL in build_state paths

[Ido Schimmel <idosch@nvidia.com>]

On Fri, Nov 06, 2026 at 12:40:55PM +0100, Max Boone wrote:
> Since 3d25eabbbf67 ("ip_tunnel: Add dst_cache support in lwtunnel_state
> of ip tunnel"), build_state runs outside RCU and all its callers can
> sleep, but paths still use GFP_ATOMIC.
> 
> For dst caches, that limits allocation to populated percpu pages. A
> large burst of SRv6 nexthops outruns the refill worker and will fail
> with -ENOMEM despite there being free memory:
> 
>   percpu: allocation failed, size=40 align=8 atomic=1, atomic alloc
> failed, no space left
> 
> Batching 20000 RTM_NEWNEXTHOP under QEMU gave 10-109 failures per run
> before, none after.
> 
> This commit converts the allocations reached through ops->build_state()
> and are called after the rcu_read_unlock():
> 
>   - ila_build_state(), ioam6_build_state(), rpl_build_state(),
>     seg6_build_state(): build_state ops
>   - bpf_parse_prog(): only called from bpf_build_state()
>   - lwtunnel_state_alloc(): only called from build_state ops (also
>     affects mpls, ip/ip6 tunnel, xfrmi and seg6_local)
> 
> Signed-off-by: Max Boone <mboone@akamai.com>
> Reviewed-by: Debabrata Banerjee <dbanerje@akamai.com>
> Reviewed-by: Nick Hudson <nhudson@akamai.com>

Reviewed-by: Ido Schimmel <idosch@nvidia.com>