Re: [PATCH 4/4] ipvs: reduce kernel stack usage

[Willem de Bruijn]

On Fri, Jun 28, 2019 at 8:40 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> With the new CONFIG_GCC_PLUGIN_STRUCTLEAK_BYREF_ALL option, the stack
> usage in the ipvs debug output grows because each instance of
> IP_VS_DBG_BUF() now has its own buffer of 160 bytes that add up
> rather than reusing the stack slots:
>
> net/netfilter/ipvs/ip_vs_core.c: In function 'ip_vs_sched_persist':
> net/netfilter/ipvs/ip_vs_core.c:427:1: error: the frame size of 1052 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> net/netfilter/ipvs/ip_vs_core.c: In function 'ip_vs_new_conn_out':
> net/netfilter/ipvs/ip_vs_core.c:1231:1: error: the frame size of 1048 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> net/netfilter/ipvs/ip_vs_ftp.c: In function 'ip_vs_ftp_out':
> net/netfilter/ipvs/ip_vs_ftp.c:397:1: error: the frame size of 1104 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
> net/netfilter/ipvs/ip_vs_ftp.c: In function 'ip_vs_ftp_in':
> net/netfilter/ipvs/ip_vs_ftp.c:555:1: error: the frame size of 1200 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]
>
> Since printk() already has a way to print IPv4/IPv6 addresses using
> the %pIS format string, use that instead,

since these are sockaddr_in and sockaddr_in6, should that have the 'n'
specifier to denote network byteorder?

> combined with a macro that
> creates a local sockaddr structure on the stack. These will still
> add up, but the stack frames are now under 200 bytes.

would it make sense to just define a helper function that takes const
char * level and msg strings and up to three struct nf_inet_addr* and
do the conversion in there? No need for macros and no state on the
stack outside error paths at all.

>
> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> ---
> I'm not sure this actually does what I think it does. Someone
> needs to verify that we correctly print the addresses here.
> I've also only added three files that caused the warning messages
> to be reported. There are still a lot of other instances of
> IP_VS_DBG_BUF() that could be converted the same way after the
> basic idea is confirmed.
> ---
>  include/net/ip_vs.h             | 71 +++++++++++++++++++--------------
>  net/netfilter/ipvs/ip_vs_core.c | 44 ++++++++++----------
>  net/netfilter/ipvs/ip_vs_ftp.c  | 20 +++++-----
>  3 files changed, 72 insertions(+), 63 deletions(-)
>
> diff --git a/include/net/ip_vs.h b/include/net/ip_vs.h
> index 3759167f91f5..3dfbeef67be6 100644
> --- a/include/net/ip_vs.h
> +++ b/include/net/ip_vs.h
> @@ -227,6 +227,16 @@ static inline const char *ip_vs_dbg_addr(int af, char *buf, size_t buf_len,
>                        sizeof(ip_vs_dbg_buf), addr,                     \
>                        &ip_vs_dbg_idx)
>
> +#define IP_VS_DBG_SOCKADDR4(fam, addr, port)                           \
> +       (struct sockaddr*)&(struct sockaddr_in)                         \
> +       { .sin_family = (fam), .sin_addr = (addr)->in, .sin_port = (port) }

might as well set .sin_family = AF_INET here and AF_INET6 below?

> +#define IP_VS_DBG_SOCKADDR6(fam, addr, port)                           \
> +       (struct sockaddr*)&(struct sockaddr_in6) \
> +       { .sin6_family = (fam), .sin6_addr = (addr)->in6, .sin6_port = (port) }