Re: [PATCH printk 3/5] printk: use buffer pool for sprint buffers

[Petr Mladek]

On Thu 2020-09-24 11:49:06, Rasmus Villemoes wrote:
> On 24/09/2020 10.53, Sergey Senozhatsky wrote:
> > On (20/09/24 10:45), Petr Mladek wrote:
> >> On Thu 2020-09-24 14:40:58, Sergey Senozhatsky wrote:
> >>> On (20/09/23 17:11), Petr Mladek wrote:
> >>>>
> >>>> AFAIK, there is one catch. We need to use va_copy() around
> >>>> the 1st call because va_format can be proceed only once.
> >>>>
> >>>
> >>> Current printk() should be good enough for reporting, say, "Kernel
> >>> stack overflow" errors. Is extra pressure that va_copy() adds something
> >>> that we need to consider?
> >>
> >> The thing is that vsprintf() traverses the arguments using va_arg().
> >> It modifies internal values so that the next va_arg() will read
> >> the next value.
> > 
> > Yes, I understand the purpose of va_copy(). I'm asking if we are
> > always on the safe side doing va_copy for every printk (+ potential
> > recursive va_copy-s).
> 
> va_copy doesn't really add any extra stack use worth talking about. When
> ABI says all arguments are passed on stack, va_list is just a pointer
> into the arguments, and va_copy merely copies that pointer. When some
> arguments are passed in register, the function prologue sets up a
> "register save area" where those registers are stashed, and va_list then
> contains two pointers: one to the reg save area, one to the place in the
> stack where the rest of the arguments are, plus a bit of control
> information on how many of the register arguments have been consumed so
> far (and that control info is the only reason one must "reset" a
> va_list, or rather use a copy that was made before consumption started).
> In either case, an extra va_list doesn't cost more than 24 bytes or so
> of stack - even if printk() was called with 100 arguments, those
> 90-100ish arguments are only stored once on the stack.

Thanks a lot for the detailed information. It is good to know that
va_copy() does not cost that much.

Best Regards,
Petr