On Mon, Feb 26, 2024, at 08:09, Xi Ruoyao wrote:
> On Mon, 2024-02-26 at 07:56 +0100, Arnd Bergmann wrote:
>> On Mon, Feb 26, 2024, at 07:03, Icenowy Zheng wrote:
>> > 在 2024-02-25星期日的 15:32 +0800，Xi Ruoyao写道：
>> > > On Sun, 2024-02-25 at 14:51 +0800, Icenowy Zheng wrote:
>> > > > My idea is this problem needs syscalls to be designed with deep
>> > > > argument inspection in mind; syscalls before this should be
>> > > > considered
>> > > > as historical error and get fixed by resotring old syscalls.
>> > >
>> > > I'd not consider fstat an error as using statx for fstat has a
>> > > performance impact (severe for some workflows), and Linus has
>> > > concluded
>> >
>> > Sorry for clearance, I mean statx is an error in ABI design, not fstat.
>
> I'm wondering why we decided to use AT_EMPTY_PATH/"" instead of
> "AT_NULL_PATH"/nullptr in the first place?
Not sure, but it's hard to change now since the libc
implementation won't easily know whether using the NULL
path is safe on a given kernel. It could check the kernel
version number, but that adds another bit of complexity in
the fast path and doesn't work on old kernels with the
feature backported.
> But it's not irrational to pass a path to syscall, as long as we still
> have the concept of file system (maybe in 2371 or some year we'll use a
> 128-bit UUID instead of path).
>
>> The problem I see with the 'use use fstat' approach is that this
>> does not work on 32-bit architectures, unless we define a new
>> fstatat64_time64() syscall, which is one of the things that statx()
>
> "fstat64_time64". Using statx for fstatat should be just fine.
Right. It does feel wrong to have only an fstat() variant but not
fstatat() if we go there.
> Or maybe we can just introduce a new AT_something to make statx
> completely ignore pathname but behave like AT_EMPTY_PATH + "".
I think this is better than going back to fstat64_time64(), but
it's still not great because
- all the reserved flags on statx() are by definition incompatible
with existing kernels that return -EINVAL for any flag they do
not recognize.
- you still need to convince libc developers to actually use
the flag despite the backwards compatibility problem, either
with a fallback to the current behavior or a version check.
Using the NULL path as a fallback would solve the problem with
seccomp, but it would not make the normal case any faster.
>> was trying to avoid.
>
> Oops. I thought "newstat" should be using 64-bit time but it seems the
> "new" is not what I'd expected... The "new" actually means "newer than
> Linux 0.9"! :(
>
> Let's not use "new" in future syscall names...
Right, we definitely can't ever succeed. On some architectures
we even had "oldstat" and "stat" before "newstat" and "stat64",
and on some architectures we mix them up. E.g. x86_64 has fstat()
and fstatat64() with the same structure but doesn't define
__NR_newfstat. On mips64, there is a 'newstat' but it has 32-bit
timestamps unlike all other 64-bit architectures.
statx() was intended to solve these problems once and for all,
and it appears that we have failed again.
Arnd
