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

[Laurent Vivier]

Le 02/07/2016 à 22:12, Peter Maydell a écrit :
> 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:

Where is defined this __ARCH_WANT_SYS_OLD_SELECT?

>  arm, m68k, mn10300, x86
> and no others use sys_old_select.

You're right, NR_select is sys_ni_syscall for:

microblaze, mips32, sh4, ppc64

arch/microblaze/kernel/syscall_table.S: .long sys_ni_syscall            /*
old_select */
arch/mips/kernel/scall32-o32.S: PTR     sys_ni_syscall                  /* 
old_select */
arch/sh/kernel/syscalls_32.S:   .long sys_ni_syscall    /* sys_oldselect */
arch/pwoerpc/include/asm/systbl.h:SYSX(sys_ni_syscall,sys_ni_syscall,ppc_select)

but I have supposed that it was set to sys_old_select for older kernel.
[but in 1.3.48, it is already sys_ni_syscall for mips... so we must
really manage that as ENOSYS)

In x86, old_select is used for the 32bit version, not for the 64bit:

entry/syscalls/syscall_32.tbl

82      i386    select                  sys_old_select
compat_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

They use:

kernel/sys.c:

#undef __SYSCALL
#define __SYSCALL(nr, call) [nr] = (call),

void *sys_call_table[__NR_syscalls] = {
#include <asm/unistd.h>
};

It's not very clear, but I think they use NR_select with sys_select:

include/uapi/asm-generic/unistd.h

#define __ARCH_WANT_SYS_SELECT
__SYSCALL(__NR_select, sys_select)

> (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 they really exist (from asm-generic/unistd.h)

> 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).

I agree.

Thanks,
Laurent