[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 16 Feb 2022 18:52:43 -0700
From: David Ahern <dsahern@...il.com>
To: Guillaume Nault <gnault@...hat.com>,
Jakub Kicinski <kuba@...nel.org>,
Jamal Hadi Salim <jhs@...atatu.com>,
Cong Wang <xiyou.wangcong@...il.com>
Cc: stephen@...workplumber.org, netdev@...r.kernel.org
Subject: Re: [RFC iproute2] tos: interpret ToS in natural numeral system
On 2/16/22 3:23 PM, Guillaume Nault wrote:
> On Wed, Feb 16, 2022 at 11:42:05AM -0800, Jakub Kicinski wrote:
>> Silently forcing a base numeral system is very painful for users.
>> ip currently interprets tos 10 as 0x10. Imagine user's bash script
>> does:
>>
>> .. tos $((TOS * 2)) ..
>>
>> or any numerical operation on the ToS.
>>
>> This patch breaks existing scripts if they expect 10 to be 0x10.
>
> I agree that we shouldn't have forced base 16 in the first place.
> But after so many years I find it a bit dangerous to change that.
I agree. In this case the change in behavior will not be very obvious
and could lead to confusion. I think this is something we have to live with.
>
> What about just printing a warning when the value isn't prefixed with
> '0x'? Something like (completely untested):
>
> @@ -535,6 +535,12 @@ int rtnl_dsfield_a2n(__u32 *id, const char *arg)
> if (!end || end == arg || *end || res > 255)
> return -1;
> *id = res;
> +
> + if (strncmp("0x", arg, 2))
> + fprintf(stderr,
> + "Warning: dsfield and tos parameters are interpreted as hexadecimal values\n"
> + "Use 'dsfield 0x%02x' to avoid this message\n", res);
> +
> return 0;
> }
That seems reasonable to me to let users know of this behavior.
>
>
>> Signed-off-by: Jakub Kicinski <kuba@...nel.org>
>> --
>> I get the feeling this was discussed in the past.
>>
>> Also there's more:
>>
>> devlink/devlink.c: val = strtoull(str, &endptr, 10);
>> devlink/devlink.c: val = strtoul(str, &endptr, 10);
>> devlink/devlink.c: val = strtoul(str, &endptr, 10);
>> devlink/devlink.c: val = strtoul(str, &endptr, 10);
I think the devlink ones all check out. I think those functions can be
updated to use get_uNN from lib/utils.c.
>> lib/utils.c: res = strtoul(arg, &ptr, base);
This is get_unsigned and those users outside of tc look ok.
>> lib/utils.c: n = strtoul(cp, &endp, 16);
ipv6 address conversion which is hex based.
>> lib/utils.c: tmp = strtoul(tmpstr, &endptr, 16);
>> lib/utils.c: tmp = strtoul(arg + i * 3, &endptr, 16);
both of these are valid hex conversions
>> misc/lnstat_util.c: unsigned long f = strtoul(ptr, &ptr, 16);
>> tc/f_u32.c: htid = strtoul(str, &tmp, 16);
>> tc/f_u32.c: hash = strtoul(str, &tmp, 16);
>> tc/f_u32.c: nodeid = strtoul(str, &tmp, 16);
>> tc/tc_util.c: maj = strtoul(str, &p, 16);
>> tc/tc_util.c: maj = strtoul(str, &p, 16);
>> tc/tc_util.c: min = strtoul(str, &p, 16);
>
> After a very quick look, many of these seem to make sense though. For
> example lnstat_util.c parses output from /proc, hexstring_a2n() is used
> by ipmacsec.c to parse crypto keys, etc.
>
> But I agree that we should probably audit the strtol() (and variants)
> that don't use base 0.
Added Jamal and Cong for the tc references. Hopefully those are
documented to be hex values.
Powered by blists - more mailing lists