On Tue, Jun 25, 2024 at 3:24 PM Xi Ruoyao <xry111@xxxxxxxxxxx> wrote:
>
> On Tue, 2024-06-25 at 13:00 +0200, Mateusz Guzik wrote:
> > + if (flags == AT_EMPTY_PATH && vfs_empty_path(dfd, filename))
>
> Could it be
>
> if ((flags & AT_EMPTY_PATH) && vfs_empty_path(dfd, filename))
>
> instead?
>
> When fstatat is implemented with statx AT_NO_AUTOMOUNT is needed, or at
> least Glibc developers think it's needed:
>
> #if FSTATAT_USE_STATX
>
> static inline int
> fstatat64_time64_statx (int fd, const char *file, struct __stat64_t64 *buf,
> int flag)
> {
> /* 32-bit kABI with default 64-bit time_t, e.g. arc, riscv32. Also
> 64-bit time_t support is done through statx syscall. */
> struct statx tmp;
> int r = INTERNAL_SYSCALL_CALL (statx, fd, file, AT_NO_AUTOMOUNT | flag,
> STATX_BASIC_STATS, &tmp);
>
> so "flags == AT_EMPTY_PATH" won't be true if Glibc implements fstatat
> and fstat via statx (on LoongArch and 32-bit platforms).
>
> I was just surprised when I saw a 100%+ improve for statx("",
> AT_EMPTY_PATH) but not stat on the Loongson machine.
>
It can't be like that specifically because we still need to catch
bogus AT flags.
I'm going to poke a little bit and send a v2, thanks.
--
Mateusz Guzik <mjguzik gmail.com>
