[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH 02/12 v9] linux-user: tilegx: Add target feature
From: |
Peter Maydell |
Subject: |
Re: [Qemu-devel] [PATCH 02/12 v9] linux-user: tilegx: Add target features support within qemu |
Date: |
Fri, 10 Apr 2015 22:51:57 +0100 |
On 10 April 2015 at 21:41, Chen Gang <address@hidden> wrote:
> On 4/10/15 05:31, Peter Maydell wrote:
>> On 27 March 2015 at 10:49, Chen Gang <address@hidden> wrote:
>>> +typedef struct target_sigaltstack {
>>> + abi_ulong ss_sp;
>>> + abi_ulong ss_size;
>>> + abi_long ss_flags;
>>> +} target_stack_t;
>>
>> Where does this come from? It doesn't match the kernel's
>> generic-headers struct layout.
>>
>
> Oh, sorry, originally, I guess, I only copied it from microblaze, did
> not check kernel.
These structures are all user-guest-facing ABI, so they must
match the kernel's structures for your target architecture.
> I shall use generic-headers which tilegx will use (the result will like
> alpha has done):
>
> typedef struct target_sigaltstack {
> abi_ulong ss_sp;
> int32_t ss_flags;
> int32_t dummy;
> abi_ulong ss_size;
> } target_stack_t;
This doesn't match the kernel either.
http://lxr.free-electrons.com/source/include/uapi/asm-generic/signal.h#L111
You have a pointer, an int and a size_t, so you want
abi_ulong ss_sp;
abi_int ss_flags;
abi_ulong ss_size;
like aarch64.
>
> [...]
>>> +
>>> +struct target_ipc_perm {
>>> + abi_int __key; /* Key. */
>>> + abi_uint uid; /* Owner's user ID. */
>>> + abi_uint gid; /* Owner's group ID. */
>>> + abi_uint cuid; /* Creator's user ID. */
>>> + abi_uint cgid; /* Creator's group ID. */
>>> + abi_uint mode; /* Read/write permission. */
>>> + abi_ushort __seq; /* Sequence number. */
>>> + abi_ushort __pad2;
>>> + abi_ulong __unused1;
>>> + abi_ulong __unused2;
>>> +};
>>
>> Again, doesn't seem to match kernel?
>>
>
> For me, it matches kernel. mode is abi_uint (__kernel_mode_t is 32-bit).
I'm looking at
http://lxr.free-electrons.com/source/include/uapi/linux/ipc.h#L9
which doesn't have that padding and unused fields at the end.
However the ipc structs are pretty confusing so maybe that's
the wrong one -- which one are you looking at?
-- PMM