Re: [PATCH] vsprintf: drop comment claiming %n is ignored

[Joe Perches <joe@perches.com>]

On Fri, 2013-09-13 at 16:03 -0700, Kees Cook wrote:
> On Fri, Sep 13, 2013 at 3:27 PM, Joe Perches <joe@perches.com> wrote:
> > On Fri, 2013-09-13 at 15:53 -0400, George Spelvin wrote:
> >> > Maybe WARN_ONCE so it's easier to emit the format too.
> >>
> >> Good idea.  And, if it's not too much trouble, a comment explaining
> >> why it's deliberately omitted so the issue doesn't arise again.
> >
> > Before any of the %n uses could be removed, I believe seq_printf
> > could to be converted to return void and have a another mechanism
> > to determine if any error occurred and the length of the output of
> > seq_printf.
> >
> > I've done a preliminary conversion of seq_printf and seq_vprintf
> > to return void and added last_ret and last_len to struct seq_file.
> >
> > If that's applied, it's trivial to convert vsnprintf to skip %n.
> >
> > Anyone have an opinion of a different conversion mechanism?
> 
> Maybe I missed this somewhere in the thread, but I'm not sure I
> understand the move to "void".

Hi Kees.

Al Viro suggested just fixing the misuses.
https://lkml.org/lkml/2013/9/11/801

David Laight suggested converting to void.
https://lkml.org/lkml/2013/9/12/113

>  Here's what I see, please correct me:
> 
> 1- seq_printf currently returns success/failure
> 2- some callers of seq_printf (correctly) use the return value as
> success/failure indication
> 3- some callers of seq_printf (incorrectly) use the return value as a
> length indication
> 4- both success/failure and length are important outputs from seq_printf
> 5- we need a way to access the length written during the call

Right and I agree with all of that.

> 6- want to minimize impact on the code base

Maybe not.

> Due to 1 and 2, it seems like there's no sense in changing the return
> value to void. Success/failure is already returned, and there are
> users of it. No sense changing them.

Except that future code might expect len instead of bool.
I did and would again.

So, I really don't care much and I don't really want to
paper over existing misuses, so I choose to fix them all.

> The normal way to handle multiple return values (4 and 5) is to add a
> pointer argument. For example: seq_printf(s, &len, fmt, args...) where
> len can be NULL. But this runs against 6.
> 
> Due to 6, to solve 4 and 5, usually macro or inline tricks are used,
> for example:
> __printf(3, 4) int seq_printf_len(struct seq_file *, size_t *len, ...);
> #define seq_printf(s, fmt, ...) seq_printf_len(s, NULL, fmt, ##__VA_ARGS__)

I'd rather fix the code that's defective than
paper over the defects with macro tricks.

> With this, solving 3 becomes possible (your void patch has already
> detected all the users of the return value, so we can sort out which
> expect length and which expect success/failure),

I (believe I) did that.

> and lets us actually
> remove the %n uses trivially too.

Soon I hope.

Anyone else have an opinion?