Re: [PATCH] printf: add support for printing symbolic error codes

[Rasmus Villemoes]

On 04/09/2019 18.28, Uwe Kleine-KÃnig wrote:
> On 9/4/19 6:19 PM, Andy Shevchenko wrote:
>> On Sat, Aug 31, 2019 at 12:48 AM Rasmus Villemoes
>> <linux@xxxxxxxxxxxxxxxxxx> wrote:
>>>
>>
>>> +/*
>>> + * Ensure these tables to not accidentally become gigantic if some
>>> + * huge errno makes it in. On most architectures, the first table will
>>> + * only have about 140 entries, but mips and parisc have more sparsely
>>> + * allocated errnos (with EHWPOISON = 257 on parisc, and EDQUOT = 1133
>>> + * on mips), so this wastes a bit of space on those - though we
>>> + * special case the EDQUOT case.
>>> + */
>>> +#define E(err) [err + BUILD_BUG_ON_ZERO(err <= 0 || err > 300)] = #err
>>
>> Hmm... Perhaps better to define the upper boundary with something like
>>
>> #define __E_POSIX_UPPER_BOUNDARY 300 // name sucks, I know
>>
>>> +#define E(err) [err - 512 + BUILD_BUG_ON_ZERO(err < 512 || err > 550)] = #err
>>
>> Similar to 550?
> 
> I'd not add "POSIX" in the name. Given that the arrays are called
> codes_0 and codes_512 I don't think using plain numbers hurts much and
> choosing a good name is hard, so I suggest to keep the explicit numbers.

I agree, adding random macro names for these essentially arbitrary (and
one-time use) numbers doesn't make sense. Remember that the sizing of
the arrays is done automatically by gcc. I suppose an alternative is to
drop the BUILD_BUG_ON_ZEROs from the E() defines and then just have some
static_assert(ARRAY_SIZE(codes_0) < 300) - but the advantage of the
above is that one gets to know _which_ E* has a huge value (that is how
I caught EDQUOT on MIPS).

>>> +const char *errcode(int err)
>>> +{
>>> +       /* Might as well accept both -EIO and EIO. */
>>> +       if (err < 0)
>>> +               err = -err;
>>> +       if (err <= 0) /* INT_MIN or 0 */
>>> +               return NULL;
>>> +       if (err < ARRAY_SIZE(codes_0))
>>> +               return codes_0[err];
>>> +       if (err >= 512 && err - 512 < ARRAY_SIZE(codes_512))
>>> +               return codes_512[err - 512];
>>> +       /* But why? */
>>> +       if (IS_ENABLED(CONFIG_MIPS) && err == EDQUOT) /* 1133 */
>>> +               return "EDQUOT";
>>
>> Another possibility is to initialize the errors at run time with radix tree.
> 
> The idea was to save space. But when using a radix tree this has
> overhead compared to the lists here, and you still need a map for
> error-code -> error-name to initialize the radix tree. Also a lookup is
> slower than with the idea implemented here. So it's bigger, slower and
> more complicated ... I don't think we should do that.

Yes, a radix tree is unlikely to end up saving space at all.
Moreover, any initialization at run-time means there's some window where
we don't have them, and printk() should work as early as possible (and I
really don't want to add any kind of synchronization "are we initialized
yet", just see what that did to the pointer hashing). So I'll stick with
the arrays.

>>> @@ -2111,6 +2112,31 @@ static noinline_for_stack
>>>  char *pointer(const char *fmt, char *buf, char *end, void *ptr,
>>>               struct printf_spec spec)
>>>  {
>>> +       /* %px means the user explicitly wanted the pointer formatted as a hex value. */
>>> +       if (*fmt == 'x')
>>> +               return pointer_string(buf, end, ptr, spec);
>>
>> But instead of breaking switch case apart can we use...
>>
>>> +
>>> +       /* If it's an ERR_PTR, try to print its symbolic representation. */
>>> +       if (IS_ERR(ptr)) {
>>
>> ...  if (IS_ERR() && *fmt != 'x') {
>> here?

This makes sense, I think I'll do it that way. Thanks.

Rasmus