Re: [PATCH v2] x86: fix ktermios-termio conversion
From: Ingo Molnar
Date: Tue May 12 2009 - 05:10:37 EST
* Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Tuesday 12 May 2009, Ingo Molnar wrote:
> > Hm, that looks a bit ugly and it also adds 150 bytes of bloat to the
> > kernel:
> >
> > drivers/char/tty_ioctl.o:
> >
> > text data bss dec hex filename
> > 4704 0 0 4704 1260 tty_ioctl.o.before
> > 4841 0 0 4841 12e9 tty_ioctl.o.after
>
> It's easy to go back to a macro (or replace it by another inline
> function if you prefer that stylistically) to make it look nicer.
no, we shouldnt go back to a macro.
> Regarding the bloat, I'm still looking for a better version. The
> code I posted is basically what is in asm-generic/termios.h
> traditionally, so at least I'm reasonably confident that it is
> correct, while the x86 version currently fails to return -EFAULT
> on incorrect pointers and misses c_line changes.
>
> I now tried this version, which theoretically should be fairly
> compact on most architectures:
>
> static inline int set_low_termios_bits(unsigned int *termios,
> const short __user *termio)
> {
> unsigned short tmp;
> int ret;
>
> ret = __get_user(tmp, termio);
> *termios = (0xffff0000 & *termios) | tmp;
>
> return ret;
> }
>
> /*
> * Translate a "termio" structure into a "termios". Ugh.
> */
> static inline int user_termio_to_kernel_termios(struct ktermios *termios,
> const struct termio __user *termio)
> {
> if (access_ok(VERIFY_READ, termio, sizeof (*termio)))
> return -EFAULT;
>
> return set_low_termios_bits(&termios->c_iflag, &termio->c_iflag) |
> set_low_termios_bits(&termios->c_oflag, &termio->c_oflag) |
> set_low_termios_bits(&termios->c_cflag, &termio->c_cflag) |
> set_low_termios_bits(&termios->c_lflag, &termio->c_lflag) |
> __get_user(termios->c_line, &termio->c_line) |
> (__copy_from_user(termios->c_cc, termio->c_cc, NCC)
> ? -EFAULT : 0);
> }
>
> /*
> * Translate a "termios" structure into a "termio". Ugh.
> */
> static inline int kernel_termios_to_user_termio(struct termio __user *termio,
> struct ktermios *termios)
> {
> if (access_ok(VERIFY_WRITE, termio, sizeof (*termio)))
> return -EFAULT;
>
> return __put_user(termios->c_iflag, &termio->c_iflag) |
> __put_user(termios->c_oflag, &termio->c_oflag) |
> __put_user(termios->c_cflag, &termio->c_cflag) |
> __put_user(termios->c_lflag, &termio->c_lflag) |
> __put_user(termios->c_line, &termio->c_line) |
> (__copy_to_user(termio->c_cc, termios->c_cc, NCC)
> ? -EFAULT : 0);
> }
>
> Unfortunately, this is even more code on x86, where get_user/put_user
> is out-of-line, while __get_user/__put_user is inline and produces
> fixup code.
> The best I could come up with is somewhat slower but also shorter:
>
> static inline int set_low_termios_bits(unsigned int *termios,
> const short __user *termio)
> {
> unsigned short tmp;
> int ret;
>
> ret = get_user(tmp, termio);
> *termios = (0xffff0000 & *termios) | tmp;
>
> return ret;
> }
>
> /*
> * Translate a "termio" structure into a "termios". Ugh.
> */
> static inline int user_termio_to_kernel_termios(struct ktermios *termios,
> const struct termio __user *termio)
> {
> return (set_low_termios_bits(&termios->c_iflag, &termio->c_iflag) ||
> set_low_termios_bits(&termios->c_oflag, &termio->c_oflag) ||
> set_low_termios_bits(&termios->c_cflag, &termio->c_cflag) ||
> set_low_termios_bits(&termios->c_lflag, &termio->c_lflag) ||
> get_user(termios->c_line, &termio->c_line) ||
> copy_from_user(termios->c_cc, termio->c_cc, NCC))
> ? -EFAULT : 0;
> }
>
> /*
> * Translate a "termios" structure into a "termio". Ugh.
> */
> static inline int kernel_termios_to_user_termio(struct termio __user *termio,
> struct ktermios *termios)
> {
> return put_user(termios->c_iflag, &termio->c_iflag) ||
> put_user(termios->c_oflag, &termio->c_oflag) ||
> put_user(termios->c_cflag, &termio->c_cflag) ||
> put_user(termios->c_lflag, &termio->c_lflag) ||
> put_user(termios->c_line, &termio->c_line) ||
> (copy_to_user(termio->c_cc, termios->c_cc, NCC))
> ? -EFAULT : 0;
> }
>
> If you think that looks better, I can send another patch. I have
> not tested this code functionally though.
hm, on x86 you could use get_user_try / get_user_ex() /
get_user_catch() approach to linearize the dependencies and reduce
register pressure.
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/