Re: [Qemu-trivial] [Qemu-devel] [PATCH] linux-user: fix signal() syscall on x86_64

[Peter Maydell]

On 2 July 2016 at 17:41, Laurent Vivier <address@hidden> wrote:
> Sadly, this can't work:
>
> sparc/sparc64/cris use sys_select for NR_select AND NR_newselect.

> Not sure all is correct, but it's what I've found:
>
>             | __NR_select    | __NR__newselect
> ------------+----------------+-----------------+
> arm         | sys_old_select | sys_select      |
> ------------+----------------+-----------------+
> aarch64     | sys_select     |        -        |
> ------------+----------------+-----------------+
> alpha       | sys_select     |        -        |
> ------------+----------------+-----------------+
> cris        | sys_select     | sys_select      |
> ------------+----------------+-----------------+
> m68k        | sys_old_select | sys_select      |
> ------------+----------------+-----------------+
> microblaze  | sys_old_select | sys_select      |
> ------------+----------------+-----------------+
> mips        | sys_old_select | sys_select      |
> ------------+----------------+-----------------+
> mips64      | sys_select     |        -        |
> ------------+----------------+-----------------+
> openrisc    | sys_select     |        -        |
> ------------+----------------+-----------------+
> ppc         | sys_old_select | sys_select      |
> ------------+----------------+-----------------+
> s390x       | sys_select     |        -        |
> ------------+----------------+-----------------+
> sh4         | sys_old_select | sys_select      |
> ------------+----------------+-----------------+
> sparc       | sys_select     | sys_select      |
> ------------+----------------+-----------------+
> sparc64     | sys_select     | sys_select      |
> ------------+----------------+-----------------+
> tilegx      | sys_select     |        -        |
> ------------+----------------+-----------------+
> unicore32   | sys_select     |        -        |
> ------------+----------------+-----------------+
> x86_64      | sys_select     |        -        |
> ------------+----------------+-----------------+
> i386        | sys_old_select | sys_select      |
> ------------+----------------+-----------------+

Hmm. Looking at current Linux git master, I get
slightly different results. The only architectures which
define __ARCH_WANT_SYS_OLD_SELECT are:
 arm, m68k, mn10300, x86
and no others use sys_old_select.

So I think we have the following behaviours:

(1) Define neither NR_select nor NR__newselect
 (and use pselect6 syscall for select):
 aarch64, openrisc, tilegx, unicore32, presumably any future arch

(2) only define NR__newselect, it is new select:
 mips, mips64, sh, s390

(3) Only define NR_select, want that to be new select:
 alpha, x86_64, s390x

(4) NR__newselect is new select, NR_select is old_select:
 i386, m68k, arm if kernel is not CONFIG_AEABI

(5) NR__newselect is new select, NR_select is defined but
 if called returns ENOSYS:
 microblaze, arm if CONFIG_AEABI, ppc64

(6) NR__newselect is new select, NR_select is a bonkers custom
 thing that tries to autodetect the calling convention:
http://lxr.free-electrons.com/source/arch/powerpc/kernel/syscalls.c#L86
 ppc32 (but only native 32-bit; 32-bit compat support
 on a ppc64 kernel is category 5, so I vote for ignoring
 this weirdness and calling ppc category 5)

(7) NR_select and NR__newselect are different numbers
 but both are new select:
 cris, sparc, sparc64

which is a pretty confusing mess, but I think it equates to:
(0) if defined, NR__newselect is always new select
(1) if NR_select is defined, the choices are:
 (a) NR_select is old_select:
   i386, m68k, arm
 (b) NR_select is defined but should ENOSYS:
   microblaze, ppc
 (c) NR_select defined and is new select:
   everything else (alpha, x86-64, s390x, cris, sparc, sparc64)

and I think we should handle that by having the code in syscall.c
be something like:

#ifdef TARGET_NR_select
    case TARGET_NR_select:
#if defined(TARGET_WANT_NI_OLD_SELECT)
        /* some architectures used to have old_select here
         * but now ENOSYS it.
         */
        ret = -TARGET_ENOSYS;
        break;
#elif defined(TARGET_WANT_OLD_SYS_SELECT)
        /* code for old select here; maybe factored out to
         * its own function: ret = do_old_select() ?
         */
#else
        /* select is new style select */
        ret = do_select(...);
#endif
#endif

where TARGET_WANT_NI_OLD_SELECT and
TARGET_WANT_OLD_SYS_SELECT are #defined in
linux-user/$(ARCH)/target_syscall.h by those
architectures that need that behaviour
(microblaze, ppc for the first; i386, m68k, arm
for the second).

We could just not define TARGET_NR_select for
microblaze and ppc, of course, but that might
be confusing and easily accidentally reverted.

For openrisc, sh and tilegx we incorrectly define
a TARGET_NR_select which the kernel doesn't, so
we should delete that from our headers.

I think overall that produces a reasonable separation
of "what behaviour does my architecture want" from
the implementation of the various behaviours, and
means the default will be correct for any architectures
we add later (only the oddball legacy cases need
to request special behaviour).

thanks
-- PMM